Method, device and equipment for optimizing topology of data center network

Abstract

The application relates to a data center network-oriented topology optimization method, device and equipment. The method comprises the following steps: constructing a continuous latent space representation model based on a graph variational autoencoder. A candidate latent vector is iteratively generated by collecting a function by adopting a Bayesian optimization framework. The candidate latent vector is decoded into a specific candidate network topology, and a large language model is integrated to infer and evaluate the structural features, so that a low-fidelity score is obtained. According to the low-fidelity score, the candidate network topology meeting the preset condition is screened out and submitted to a high-fidelity network simulator for performance evaluation, so that a high-fidelity performance index is obtained. The evaluation result is fed back to the Bayesian optimization framework, and the corresponding proxy model is updated to generate the candidate latent vector of the next iteration, and the whole process optimization process is cycled until the preset termination condition is met, and the optimized network topology is output. By adopting the method, the complex network topology can be efficiently and intelligently and automatically optimized.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence network optimization technology, and in particular to a method, apparatus and equipment for optimizing the topology of data center networks. Background Technology

[0002] Currently, network topology design is evolving from static rule templates to dynamic intelligent optimization. Initially, rule-based structures were commonly used to ensure engineering standardization, but their performance was limited when facing heterogeneous services. Subsequently, methods combining heuristic algorithms (such as genetic algorithms) and simulation were used to search for better topologies, but these suffered from low search efficiency and slow convergence. In recent years, graph representation learning and Bayesian optimization have provided new approaches to continuous space topology search, and large language models have also brought new possibilities for structure understanding and generation. However, these technologies are often applied in isolation and have not yet formed a deeply integrated and efficiently collaborative complete experimental design system.

[0003] At the optimization method level, existing technologies have significant shortcomings. First, the starting point of topology search is of poor quality, lacking effective prior knowledge guidance and relying heavily on random generation or simple perturbations, resulting in a large amount of simulation budget being wasted on ineffective or low-performance structures. Second, the evaluation process is inefficient, heavily reliant on time-consuming high-fidelity simulations, and lacks a fast, low-cost structure pre-screening mechanism, making it difficult to achieve efficient exploration with limited resources.

[0004] At the system design and engineering practice level, existing solutions also have limitations. Their topology generation, optimization algorithms and performance evaluation modules are usually highly coupled, forming a closed, siloed architecture that makes it difficult to flexibly integrate new search strategies or auxiliary models (such as large language models). The system's scalability and adaptability are insufficient, which restricts the rapid iteration and practical deployment of the technology. Summary of the Invention

[0005] Therefore, it is necessary to provide a method, apparatus, and device for improving the topology optimization of data center networks in response to the above-mentioned technical problems.

[0006] A topology optimization method for data center networks, the method comprising:

[0007] The network topology of the data center is formalized as an undirected graph, and a continuous latent space representation model based on graph variational autoencoder is constructed.

[0008] In the latent space representation model, a Bayesian optimization framework is adopted to generate candidate latent vectors through iterative acquisition functions.

[0009] For each round of iteration, the candidate latent vectors generated are decoded into specific candidate network topologies, and a large language model is integrated to infer and evaluate the structural features of the candidate network topologies to obtain a low-fidelity score.

[0010] Based on the low-fidelity score, candidate network topologies that meet the preset conditions are selected and submitted to the high-fidelity network simulator for performance evaluation to obtain high-fidelity performance indicators.

[0011] The evaluation results are fed back to the Bayesian optimization framework, which updates the corresponding surrogate model to generate candidate latent vectors for the next iteration. The entire optimization process is repeated until the preset termination condition is met, and the optimized network topology is output.

[0012] A topology optimization device for data center networks, the device comprising:

[0013] The latent space model building module is used to formalize the network topology of the data center into an undirected graph and build a continuous latent space representation model based on a graph variational autoencoder.

[0014] The candidate latent vector generation module is used to generate candidate latent vectors in the latent space representation model by using a Bayesian optimization framework and iteratively generating the acquisition function.

[0015] The structure prior evaluation module is used to decode the candidate latent vectors generated in each iteration into specific candidate network topologies, and integrates a large language model to infer and evaluate the structural features of the candidate network topologies to obtain low-fidelity scores.

[0016] The high-fidelity evaluation module is used to filter candidate network topologies that meet preset conditions based on low-fidelity scores, and submit them to the high-fidelity network simulator for performance evaluation to obtain high-fidelity performance indicators.

[0017] The optimization module is used to feed the evaluation results back to the Bayesian optimization framework, update the surrogate model corresponding to the Bayesian optimization framework, generate candidate latent vectors for the next iteration, and repeat the entire optimization process until the preset termination condition is met, and output the optimized network topology.

[0018] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program performing the following steps:

[0019] The network topology of the data center is formalized as an undirected graph, and a continuous latent space representation model based on graph variational autoencoder is constructed.

[0020] In the latent space representation model, a Bayesian optimization framework is adopted to generate candidate latent vectors through iterative acquisition functions.

[0021] For each round of iteration, the candidate latent vectors generated are decoded into specific candidate network topologies, and a large language model is integrated to infer and evaluate the structural features of the candidate network topologies to obtain a low-fidelity score.

[0022] Based on the low-fidelity score, candidate network topologies that meet the preset conditions are selected and submitted to the high-fidelity network simulator for performance evaluation to obtain high-fidelity performance indicators.

[0023] The evaluation results are fed back to the Bayesian optimization framework, which updates the corresponding surrogate model to generate candidate latent vectors for the next iteration. The entire optimization process is repeated until the preset termination condition is met, and the optimized network topology is output.

[0024] The aforementioned method, apparatus, and equipment for optimizing data center network topology firstly utilize a graph variational autoencoder to map a discrete, high-dimensional network topology graph into a low-dimensional, continuous latent space. This crucial step transforms the original combinatorial explosion problem of discrete structure search into an optimization problem in a smooth, continuous space, enabling the application of efficient gradient-based search algorithms (implemented through Bayesian optimization). This fundamentally improves the directness and efficiency of the optimization process, constructing a unified simulation-driven black-box graph optimization framework. Secondly, a large language model (LLM) is introduced as a prior knowledge engine. In each optimization iteration, instead of directly feeding all generated candidate topologies into time-consuming high-fidelity simulations, the LLM first performs rapid reasoning and evaluation of their structural features, generating a low-fidelity score. This mechanism acts as an intelligent filter, capable of preemptively eliminating obviously unreasonable or inefficient designs based on the structural rationality and robustness learned by the LLM from massive network knowledge. This concentrates valuable simulation resources on the most promising candidate solutions, achieving a qualitative improvement in sample utilization efficiency. Furthermore, adaptive end-to-end decision optimization is achieved through a Bayesian optimization framework. Bayesian optimization uses a Gaussian process as a surrogate model, intelligently balancing exploration and utilization within the latent space to sequentially propose better candidate points. More importantly, the results of each high-fidelity simulation evaluation are fed back in real time to update the surrogate model, forming an automated closed loop of generation-screening-evaluation-feedback. This closed loop not only enables the search process to adaptively focus on high-performance regions, but also, by integrating the low-cost structural priors provided by LLM with the high-fidelity performance feedback provided by the simulator, allows the system to autonomously discover high-performance, novel network topologies that surpass traditional rule templates, greatly enhancing the method's versatility and applicability. In summary, this scheme, through the synergistic innovation of three core technologies—continuous latent space representation, efficient sample screening guided by LLM priors, and adaptive closed-loop learning driven by Bayesian optimization—achieves efficient, intelligent, and automated optimization of complex network topologies under strict simulation budgets. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating a topology optimization method for a data center network in one embodiment.

[0026] Figure 2 This is a flowchart illustrating a Topology-Language-Autoencoder Bayesian Optimization (TLA-BO) method for data center networks in one embodiment.

[0027] Figure 3 This is a general framework diagram for topology optimization of a data center network in one embodiment;

[0028] Figure 4 This is a diagram illustrating the expected performance of a topology optimization method for data center networks in one embodiment.

[0029] Figure 5 This is a structural block diagram of a topology optimization device for a data center network in one embodiment;

[0030] Figure 6 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0032] In one embodiment, such as Figure 1 As shown, a topology optimization method for data center networks is provided, including the following steps:

[0033] Step 102: The network topology of the data center is formalized as an undirected graph, and a continuous latent space representation model based on graph variational autoencoder is constructed.

[0034] Step 104: In the latent space representation model, a Bayesian optimization framework is adopted to generate candidate latent vectors through iterative acquisition functions.

[0035] Step 106: For each round of iteration, the candidate latent vectors generated are decoded into specific candidate network topologies, and a large language model is integrated to infer and evaluate the structural features of the candidate network topologies to obtain a low-fidelity score.

[0036] Step 108: Based on the low-fidelity score, select candidate network topologies that meet the preset conditions, submit them to the high-fidelity network simulator for performance evaluation, and obtain high-fidelity performance indicators.

[0037] Step 110: Feed the evaluation results back to the Bayesian optimization framework, update the surrogate model corresponding to the Bayesian optimization framework to generate candidate latent vectors for the next iteration, and repeat the entire optimization process until the preset termination condition is met, and output the optimized network topology.

[0038] In the aforementioned topology optimization method for data center networks, firstly, a graph variational autoencoder is used to map the discrete, high-dimensional network topology graph to a low-dimensional, continuous latent space. This crucial step transforms the original combinatorial explosion problem of discrete structure search into an optimization problem in a smooth, continuous space, enabling the application of efficient gradient-based search algorithms (implemented through Bayesian optimization). This fundamentally improves the directness and efficiency of the optimization process, constructing a unified simulation-driven black-box graph optimization framework. Secondly, a large language model (LLM) is introduced as a prior knowledge engine. In each optimization iteration, instead of directly feeding all generated candidate topologies into time-consuming high-fidelity simulations, the LLM first performs rapid reasoning and evaluation of their structural features, generating a low-fidelity score. This mechanism acts as an intelligent filter, capable of preemptively eliminating obviously unreasonable or inefficient designs based on the structural rationality and robustness learned by the LLM from massive network knowledge. This concentrates valuable simulation resources on the most promising candidate solutions, achieving a qualitative improvement in sample utilization efficiency. Furthermore, adaptive end-to-end decision optimization is achieved through a Bayesian optimization framework. Bayesian optimization uses a Gaussian process as a surrogate model, intelligently balancing exploration and utilization within the latent space to sequentially propose better candidate points. More importantly, the results of each high-fidelity simulation evaluation are fed back in real time to update the surrogate model, forming an automated closed loop of generation-screening-evaluation-feedback. This closed loop not only enables the search process to adaptively focus on high-performance regions, but also, by integrating the low-cost structural priors provided by LLM with the high-fidelity performance feedback provided by the simulator, allows the system to autonomously discover high-performance, novel network topologies that surpass traditional rule templates, greatly enhancing the method's versatility and applicability. In summary, this scheme, through the synergistic innovation of three core technologies—continuous latent space representation, efficient sample screening guided by LLM priors, and adaptive closed-loop learning driven by Bayesian optimization—achieves efficient, intelligent, and automated optimization of complex network topologies under strict simulation budgets.

[0039] In one embodiment, such as Figure 2As shown, a Topology-Language-Autoencoder Bayesian Optimization (TLA-BO) method for data center networks is presented, aiming to efficiently and automatically search for the optimal network physical connection structure in terms of key indicators such as throughput, latency, and packet loss rate. Under the constraint of a limited simulation evaluation budget, this method achieves simulation-driven, efficient, and intelligent search by fusing graph representation learning, serialized Bayesian optimization, and structural priors from a large language model. The specific implementation steps of its technical solution are as follows:

[0040] The first step is data preparation: Users need to provide a well-defined configuration file, which mainly includes: network size, device connectivity limitations, optimization objectives (such as prioritizing throughput or low latency), and the maximum number of simulations allowed for this optimization task. After the system starts, it will automatically load these parameters and prepare the internal core models, including graph neural networks for understanding network structure, an optimization engine for intelligent search, and a large language model interface for providing suggestions.

[0041] The second step is intelligent parameter validation: The model automatically checks the completeness and rationality of the input parameters. If parameters are missing, contradictory, or clearly inconsistent with common engineering sense, it won't simply report an error. Instead, it will invoke a large language model to generate easy-to-understand natural language prompts and modification suggestions to guide the user in making corrections. For example, it might suggest, "Your link density setting is too low, which may cause network connectivity issues. We recommend adjusting it according to certain principles." Only after all parameters have passed validation will the optimization task officially begin.

[0042] The third step is to start the automatic optimization task: After the parameter file passes the validity check, the system will automatically parse and encapsulate all input parameters, and then call the core "Topology-Language-Autoencoder Bayesian Optimization (TLA-BO)" computing engine of this invention to start an independent optimization thread.

[0043] The core optimization process of this invention will automatically and sequentially execute the following key steps after the task is started, forming a complete intelligent design closed loop:

[0044] 3.1 Continuous Space Mapping of Graph Variational Autoencoders

[0045] First, the network topology is formalized as a constrained black-box optimization problem. The network containing N switching nodes or servers is abstracted as an undirected graph, and an adjacency matrix is ​​used to perform the optimization. The mathematical representation is used. The optimization objective is to find a feasible solution space that satisfies a series of engineering constraints, such as physical cabling distance, number of device ports, and network connectivity. Within this framework, we find the comprehensive performance metrics that maximize the evaluation of the high-fidelity network simulator. Optimal topology This metric is typically the normalized total network throughput. Compared with the overall packet loss rate The weighted combination, i.e. .

[0046] To solve the problem directly in a discrete and huge topological combination space (of size 10 ... To address the challenge of searching within a discrete graph structure (adjacency matrix), this invention introduces a graph variational autoencoder as the core representation learning component. This component, trained on a representative topological dataset, learns a representation learning algorithm that encodes discrete graph structures (adjacency matrices) within a graph. ) is mapped to a low-dimensional, continuous vector space (i.e., the latent space, ,in encoder And a decoder capable of reconstructing back-graph structures from latent vectors. Its training is achieved by maximizing the lower bound of evidence (ELBO):

[0047] ;

[0048] in For standard Gaussian priors, This step involves balancing reconstruction accuracy with latent space regularity to determine hyperparameters. Essentially, it transforms the discrete combinatorial optimization problem into one operating within a continuous, smooth latent space. Chinese pair function This optimization problem lays the foundation for applying efficient gradient-based optimization algorithms.

[0049] 3.2 Bayesian Optimization and Serialization Decision

[0050] Intelligent sequence search is performed using a Bayesian optimization framework on a constructed continuous latent space. This is because each simulation evaluation... The cost is extremely high, so a Gaussian process is used as a surrogate model to approximate this unknown complex function. A Gaussian process is... Provide a prior probability distribution, denoted as . ,in It is a mean function. It is a covariance function (such as the Matérn kernel) used to measure the performance correlation of the topologies corresponding to two latent vectors.

[0051] Based on the historical observation dataset that has been evaluated A Gaussian process can update its posterior distribution, which not only gives the result of any new point... Performance prediction mean It also provides the variance of the prediction uncertainty. By utilizing this uncertainty, the next evaluation point is intelligently selected by defining a data acquisition function. This invention uses the desired improvement as the data acquisition function:

[0052] ;

[0053] in This represents the best performance value observed so far. The function explicitly balances "exploitation" (sampling in regions with high prediction mean) and "exploration" (sampling in regions with high uncertainty) to approximate the global optimum with as few evaluations as possible.

[0054] 3.3 Large Language Model (LLM)-Driven Structural Prior Integration and Candidate Selection

[0055] This is the core innovation of the present invention. In order to further improve the search efficiency under strict simulation budget and incorporate human network design experience, a large language model is explicitly integrated as a low-cost, knowledge-rich "structural advisor" in the Bayesian optimization generation-evaluation loop.

[0056] This integration is mainly reflected in two aspects: 1. High-quality initialization: Before optimization begins, the system not only uses random initialization but also leverages LLM (Layered Management Model). By inputting textual descriptions of the target network size, typical regularized topologies (such as FatTree), and engineering constraints (such as maximum node degree) into the LLM, it guides the LLM to generate or deduce a batch of candidate topologies that are structurally reasonable and satisfy the basic constraints. These topologies, after being encoded, constitute the initial observation points in the latent space, significantly improving the quality of the search's starting point. 2. Low-fidelity screening during iteration: In each round of Bayesian optimization iteration, when the acquisition function proposes a new latent vector... Then, it is first decoded into a specific candidate topology. Next, the system automatically extracts a series of computable structural features from the topology (such as node degree distribution, network diameter, average clustering coefficient, and the presence of critical cut points), and organizes these features along with engineering constraints into structured text prompts, which are then submitted to the LLM. Based on its domain knowledge learned from massive amounts of network design literature and technical documents, the LLM quickly infers the "structural rationality," "robustness," and "potential bottleneck risk" of the candidate topology, and outputs a quantitative score. This score serves as a low-fidelity signal, compared to a high-fidelity simulation signal that has not yet been obtained. Together they constitute a multi-fidelity decision function.

[0057] For example, definition ,in This is a decaying weighting coefficient. In practice, the system prioritizes fast filtering based on LLM scores, only including... Above the threshold or comprehensive assessment A small number of candidate topologies with high expected performance are submitted to a time-consuming network simulator for precise evaluation. This mechanism effectively avoids wasting valuable simulation resources on a large number of topologies with obviously unreasonable structures or inherently poor performance.

[0058] 3.4 High-fidelity simulation verification and closed-loop iterative update

[0059] The elite candidate topologies selected through the above screening are then fed into a high-fidelity network simulation environment (such as ns-3) and run under a preset typical service traffic model to obtain their realistic and accurate performance indicators, including end-to-end throughput, tail latency, and packet loss rate, and to calculate a high-fidelity comprehensive score. This newly obtained observation pair Added to the historical dataset middle.

[0060] Subsequently, the system uses the expanded dataset to refit or update the hyperparameters of the Gaussian process surrogate model, making its modeling of the mapping relationship from latent space to performance more accurate. The updated model is used to guide the calculation of the acquisition function in the next iteration, thereby proposing new and more promising candidate points, forming a fully automated, data-driven optimization loop. This loop continues until the pre-set maximum number of simulations (evaluation budget) is reached, ultimately outputting the one or a set of network topologies that have achieved the best performance throughout the optimization process.

[0061] In one embodiment, the network topology of the data center is formalized as an undirected graph, resulting in a network topology graph structure. This discrete network topology graph structure is then encoded as a low-dimensional and continuous latent space vector. With the goal of maximizing the lower bound of evidence, the encoder and decoder of the graph variational autoencoder are trained using a dataset containing the network topology.

[0062] ;

[0063] in, To maximize the lower bound of evidence, Let be the expected function. It is a discrete network topology graph structure. For the decoder of the graph variational autoencoder, The encoder is a variational autoencoder. For latent space, For hyperparameters, For relative entropy, The prior is a standard Gaussian. The encoder maps the network topology graph structure to the latent space. The decoder reconstructs the network topology graph structure from the latent space vectors.

[0064] In one embodiment, a Gaussian process is used as a surrogate model in the latent space representation model to model the mapping relationship between latent space vectors and network topology performance indicators:

[0065] ;

[0066] in, To simulate and evaluate the probability of latent space. It is a mean function. Let covariance function be used. This is a Gaussian process. The Gaussian process provides the performance prediction mean and uncertainty variance based on historical evaluation data. Based on the prediction mean and uncertainty variance, the expected improvement in the acquisition function value is calculated:

[0067] ;

[0068] in, To improve the value of the acquisition function, Let be the expected improvement in performance at the latent space vector z, relative to the current best performance. The best performance value observed so far is selected. The latent space vector that maximizes the desired improvement in the acquisition function value is chosen as the candidate latent vector for evaluation in the next round.

[0069] In one embodiment, the candidate latent vectors generated in each iteration are decoded into specific candidate network topologies. Structural feature parameters are extracted from these candidate network topologies, including at least one of node degree distribution, network diameter, and average clustering coefficient. The structural feature parameters and preset engineering constraints are organized into structured text prompts, which are then input into a large language model. The model's quantitative score for the candidate network topology is obtained based on the structured text prompts and serves as a low-fidelity score.

[0070] In one embodiment, the large language model is used to generate one or more initial candidate network topologies that structurally satisfy the basic constraints, based on the target network size and the engineering constraints.

[0071] The initial candidate network topology is encoded into an initial latent space vector, forming the initial observation dataset for the Bayesian optimization framework:

[0072] ;

[0073] in, For the initial observation dataset, For the first The initial latent space vectors of the initial candidate network topologies. For the first Candidate network topologies Corresponding performance evaluation metrics This is the sequence number of the initial observation sample.

[0074] In one embodiment, a comprehensive evaluation score, which is composed of the low-fidelity score and the performance index predicted by the proxy model, is compared with a preset threshold. Candidate network topologies with scores higher than the preset threshold are submitted to a high-fidelity network simulator for evaluation.

[0075] It should be understood that, although Figures 1-2 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order in which these steps are executed, and they can be performed in other orders. Figures 1-2 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0076] In one embodiment, such as Figure 5 As shown, a topology optimization device for data center networks is provided, comprising: a latent space model construction module 502, a candidate latent vector generation module 504, a structure prior evaluation module 506, a high-fidelity evaluation module 508, and an optimization module 510, wherein:

[0077] The latent space model construction module 502 is used to formalize the network topology of the data center into an undirected graph and construct a continuous latent space representation model based on a graph variational autoencoder.

[0078] The candidate latent vector generation module 504 is used to generate candidate latent vectors iteratively through the acquisition function in the latent space representation model using a Bayesian optimization framework.

[0079] The structural prior evaluation module 506 is used to decode the candidate latent vectors generated in each iteration into specific candidate network topologies, and integrates a large language model to infer and evaluate the structural features of the candidate network topologies to obtain a low-fidelity score.

[0080] The high-fidelity evaluation module 508 is used to filter candidate network topologies that meet preset conditions based on the low-fidelity score, submit them to the high-fidelity network simulator for performance evaluation, and obtain high-fidelity performance indicators.

[0081] The optimization module 510 is used to feed the evaluation results back to the Bayesian optimization framework, update the surrogate model corresponding to the Bayesian optimization framework, generate candidate latent vectors for the next iteration, and repeat the entire optimization process until the preset termination condition is met, and output the optimized network topology.

[0082] Specific limitations regarding the topology optimization device for data center networks can be found in the limitations of the topology optimization method for data center networks described above, and will not be repeated here. Each module in the aforementioned topology optimization device for data center networks can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in the computer device in hardware form, or stored in the memory of the computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0083] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 6 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a topology optimization method for data center networks. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0084] Those skilled in the art will understand that Figure 3 , Figures 5-6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0085] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to perform the following steps:

[0086] The network topology of the data center is formalized as an undirected graph, and a continuous latent space representation model based on graph variational autoencoder is constructed.

[0087] In the latent space representation model, a Bayesian optimization framework is adopted to generate candidate latent vectors through iterative acquisition functions.

[0088] For each round of iteration, the candidate latent vectors generated are decoded into specific candidate network topologies, and a large language model is integrated to infer and evaluate the structural features of the candidate network topologies to obtain a low-fidelity score.

[0089] Based on the low-fidelity score, candidate network topologies that meet the preset conditions are selected and submitted to the high-fidelity network simulator for performance evaluation to obtain high-fidelity performance indicators.

[0090] The evaluation results are fed back to the Bayesian optimization framework, which updates the corresponding surrogate model to generate candidate latent vectors for the next iteration. The entire optimization process is repeated until the preset termination condition is met, and the optimized network topology is output.

[0091] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), Synchlink, DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0092] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0093] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A topology optimization method for data center networks, characterized in that, The method includes: The network topology of the data center is formalized as an undirected graph, and a continuous latent space representation model based on graph variational autoencoder is constructed. The network topology of the data center is formalized as an undirected graph to obtain the network topology graph structure. The discrete network topology graph structure is then encoded into a low-dimensional and continuous latent space vector. With the goal of maximizing the lower bound of evidence, the encoder and decoder of the graph variational autoencoder are trained using a dataset containing network topology: in, To maximize the lower bound of evidence, Let be the expected function. It is a discrete network topology graph structure. For the decoder of the graph variational autoencoder, The encoder is a variational autoencoder. For latent space, For hyperparameters, For relative entropy, Standard Gaussian priors; The encoder is used to map the network topology graph structure to the latent space; The decoder is used to reconstruct the network topology graph structure from the latent space vectors; In the latent space representation model, a Bayesian optimization framework is adopted, and candidate latent vectors are generated iteratively through the acquisition function; For each round of iteration, the candidate latent vectors generated are decoded into specific candidate network topologies, and a large language model is integrated to infer and evaluate the structural features of the candidate network topologies to obtain a low-fidelity score. Based on the low-fidelity score, candidate network topologies that meet the preset conditions are selected and submitted to the high-fidelity network simulator for performance evaluation to obtain high-fidelity performance indicators. The evaluation results are fed back to the Bayesian optimization framework, the corresponding surrogate model of the Bayesian optimization framework is updated, and candidate latent vectors for the next iteration are generated. The entire optimization process is repeated until the preset termination condition is met, and the optimized network topology is output.

2. The method according to claim 1, characterized in that, In the latent space representation model, a Bayesian optimization framework is employed to iteratively generate candidate latent vectors through an acquisition function, including: In the latent space representation model, a Gaussian process is used as a surrogate model to model the mapping relationship from latent space vectors to network topology performance indicators: in, To simulate and evaluate the probability of latent space. It is a mean function. Let covariance function be used. It is a Gaussian process; The Gaussian process provides the performance prediction mean and uncertainty variance based on historical evaluation data; Based on the predicted mean and uncertainty variance, calculate the expected improved acquisition function value: in, To improve the value of the acquisition function, Let be the expected improvement in performance at the latent space vector z, relative to the current best performance. This represents the best performance value observed so far. The latent space vector that maximizes the desired improvement in the acquisition function value is selected as the candidate latent vector to be evaluated in the next round.

3. The method according to claim 2, characterized in that, For each iteration, the candidate latent vectors generated are decoded into specific candidate network topologies, and a large language model is integrated to infer and evaluate the structural features of the candidate network topologies to obtain low-fidelity scores, including: For each round of iteration, the candidate latent vectors generated are decoded into specific candidate network topologies. Structural feature parameters are extracted from the candidate network topologies, wherein the structural feature parameters include at least one of node degree distribution, network diameter, and average clustering coefficient. The structural feature parameters and preset engineering constraints are organized into structured text prompts and input into a large language model to obtain a quantitative score of the candidate network topology based on the structured text prompts, which is used as a low-fidelity score.

4. The method according to claim 3, characterized in that, Before the step of iteratively generating candidate latent vectors through an acquisition function using a Bayesian optimization framework in the latent space representation model, the method further includes: Using the large language model, one or more initial candidate network topologies that structurally satisfy the basic constraints are generated based on the target network size and the engineering constraints. The initial candidate network topology is encoded into an initial latent space vector, forming the initial observation dataset for the Bayesian optimization framework: in, For the initial observation dataset, For the first The initial latent space vectors of the initial candidate network topologies. For the first Candidate network topologies Corresponding performance evaluation metrics This is the sequence number of the initial observation sample.

5. The method according to claim 3, characterized in that, Based on the low-fidelity score, candidate network topologies that meet preset conditions are selected, including: The comprehensive evaluation score, which is composed of the low-fidelity score and the performance index predicted by the surrogate model, is compared with a preset threshold. Candidate network topologies with scores higher than the preset threshold are submitted to the high-fidelity network simulator for evaluation.

6. A topology optimization device for data center networks, characterized in that, The device includes: The latent space model building module is used to formalize the network topology of the data center into an undirected graph and build a continuous latent space representation model based on a graph variational autoencoder. A candidate latent vector generation module is used in the latent space representation model to iteratively generate candidate latent vectors through a acquisition function using a Bayesian optimization framework; the network topology of the data center is formalized as an undirected graph to obtain a network topology graph structure, and the discrete network topology graph structure is encoded into low-dimensional and continuous latent space vectors; with the goal of maximizing the lower bound of evidence, the encoder and decoder of the graph variational autoencoder are trained using a dataset containing the network topology. in, To maximize the lower bound of evidence, Let be the expected function. It is a discrete network topology graph structure. For the decoder of the graph variational autoencoder, The encoder is a variational autoencoder. For latent space, For hyperparameters, For relative entropy, The encoder is a standard Gaussian prior; the encoder maps the network topology graph structure to the latent space; the decoder reconstructs the network topology graph structure from the latent space vectors. The structural prior evaluation module is used to decode the candidate latent vectors generated in each iteration into specific candidate network topologies, and integrate a large language model to infer and evaluate the structural features of the candidate network topologies to obtain a low-fidelity score. The high-fidelity evaluation module is used to filter candidate network topologies that meet preset conditions based on the low-fidelity score, submit them to the high-fidelity network simulator for performance evaluation, and obtain high-fidelity performance indicators. The optimization module is used to feed the evaluation results back to the Bayesian optimization framework, update the surrogate model corresponding to the Bayesian optimization framework, generate candidate latent vectors for the next iteration, and repeat the entire optimization process until the preset termination condition is met, and output the optimized network topology.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.

Images