Artificial intelligence center lossless network networking scheme generation system, method, device and medium

Abstract

The application provides a kind of intelligent calculation center lossless network networking scheme generation system, method, equipment and medium, it is related to network networking technical field, the system includes: user interaction layer, for receiving the computing power demand information input by user, and show network networking scheme;Intelligent agent arrangement layer, which is connected with user interaction layer, for driving multiple intelligent agents to execute demand analysis, topology calculation, topology mapping and bill of materials generation in turn;Rule knowledge layer, which is connected with intelligent agent arrangement layer, for storing and providing topology design rules, for intelligent agent arrangement layer to call topology design rules when executing topology calculation;Engineering verification layer, for engineering verification on the topology parameters output by intelligent agent arrangement layer, and triggering intelligent agent arrangement layer to output network networking scheme to user interaction layer after verification passes.The application realizes the automation design of intelligent calculation center lossless network networking scheme, greatly improves the design efficiency and standardization degree of networking scheme.

Description

Technical Field

[0001] This invention relates to the field of network networking technology, and in particular to a system, method, device and medium for generating lossless network networking schemes for intelligent computing centers. Background Technology

[0002] With the rapid expansion of large-scale training and inference of artificial intelligence (AI) models, intelligent computing centers have become a core infrastructure supporting the development of the AI ​​industry. Against this backdrop, the scale of Graphics Processing Unit (GPU) clusters is continuously expanding, and high-concurrency communication modes such as AllReduce place extremely high demands on the network. Network performance directly affects computing power utilization, training convergence speed, and system stability. Currently, intelligent computing center networks generally adopt the Spine-Leaf architecture, coupled with lossless network protocols such as RoCEv2, to achieve high-bandwidth, low-latency, and zero-packet-loss data transmission.

[0003] Currently, the industry has accumulated some technological expertise in the field of intelligent computing center networks, primarily focusing on physical network architecture. To reduce latency and power consumption, a converged optoelectronic data center network architecture has been proposed, replacing traditional electrical switches with Spine layer optical switches to reduce the number of optoelectronic conversions. To reduce networking costs and power consumption, a ring-structured data center layout has been proposed, achieving short-distance interconnection by shortening device spacing. To reduce communication latency, simulated annealing algorithms are used to optimize the physical location of servers. Some research has also analyzed the physical architecture and protocol selection for lossless networks. All of these existing technologies focus on the physical network architecture level, aiming to solve the problem of which physical topology and cabling method should be adopted for lossless networks in intelligent computing centers, but they do not address the automation of network design. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a system, method, device and medium for generating lossless network topology schemes for intelligent computing centers.

[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: Firstly, the present invention provides a system for generating a lossless network topology scheme for intelligent computing centers, which adopts the following technical solution: A system for generating lossless network topology schemes for intelligent computing centers includes: The user interaction layer is used to receive computing power demand information input by users and display network topology schemes. The intelligent agent orchestration layer, which is connected to the user interaction layer, is used to drive multiple intelligent agents to sequentially execute requirement parsing, topology calculation, topology drawing and bill of materials generation. The rule knowledge layer, which is connected to the agent orchestration layer, is used to store and provide topology design rules and bill of materials generation rules, so that the agent orchestration layer can call the topology design rules when performing the topology calculation and the bill of materials generation rules when performing the bill of materials generation. The engineering verification layer, which is connected to the agent orchestration layer and the user interaction layer, is used to perform engineering verification on the topology parameters output by the agent orchestration layer, and triggers the agent orchestration layer to output the network topology scheme to the user interaction layer after the verification is passed.

[0006] The beneficial effects of this invention are as follows: Through a layered architecture design comprising a user interaction layer, an intelligent agent orchestration layer, a rule knowledge layer, and an engineering verification layer, the system achieves automated design of lossless network topology schemes for intelligent computing centers. Users only need to input computing power requirements, and the system can automatically complete requirements parsing, topology calculation, topology drawing, and bill of materials generation. The engineering verification layer automatically verifies the topology parameters to ensure the engineering compliance of the generated scheme. Compared to traditional network planning methods that rely on manual experience, this invention significantly improves the design efficiency and standardization of network topology schemes, reduces the risks of human design, and can meet the needs of rapid construction and large-scale replication of intelligent computing centers.

[0007] Based on the above technical solution, the present invention can be further improved as follows.

[0008] Furthermore, the agent orchestration layer is specifically used to drive the following multiple agents to perform corresponding operations according to the agent orchestration order: A demand parsing intelligent agent is used to parse the computing power demand information into a structured demand model; A topology design agent is used to calculate the topology parameters of the lossless network of the intelligent computing center based on the structured requirements model and the topology design rules. A topology drawing agent is used to draw a network topology diagram of the lossless network of the intelligent computing center based on the topology parameters. An intelligent agent for generating equipment lists is used to generate a bill of materials for the lossless network of the intelligent computing center based on the topology parameters and the bill of materials generation rules. The solution outputs an intelligent agent, which, after the topology parameters have passed verification, generates the network topology scheme based on the structured requirements model, the topology parameters, the network topology diagram, and the bill of materials.

[0009] The beneficial effects of adopting the above-mentioned further scheme are as follows: By refining the intelligent agent orchestration layer into five dedicated intelligent agents—requirement analysis, topology design, topology drawing, equipment list generation, and scheme output—each intelligent agent has a clear division of labor and works collaboratively, forming a complete automated pipeline from requirement input to scheme output. The requirement analysis intelligent agent automatically transforms the computing power requirement information input by the user into a structured requirement model, ensuring the accuracy of subsequent calculations; the topology design intelligent agent automatically calculates topology parameters based on the structured requirement model and topology design rules, avoiding the tediousness and errors of manual calculation; the topology drawing intelligent agent automatically generates a network topology diagram, realizing the visualization of design results; the equipment list generation intelligent agent automatically generates a bill of materials, providing a direct basis for engineering procurement; and the scheme output intelligent agent integrates all results to generate a complete network topology scheme. Through the orderly collaboration of the above five intelligent agents, the entire process of generating lossless network topology schemes for intelligent computing centers is automated, further improving design efficiency and scheme standardization.

[0010] Furthermore, the structured demand model includes the total number of GPUs, the number of GPUs and network cards per server node, network card speed, and oversold ratio; the topology parameters include the number of server nodes, the number of Leaf switches, the number of uplink and downlink ports and port speeds, and the number of Spine switches and port speeds. The topology design agent is specifically used for: The number of server nodes is obtained by calculating the ratio between the total number of GPUs and the number of GPUs in a single server node; The total number of downlink ports of the Leaf switch is obtained by multiplying the number of server nodes by the number of network interface cards (NICs) of a single server node. Calculate the ratio between the total number of downlink ports of the Leaf switch and the oversold ratio to obtain the total number of uplink ports of the Leaf switch; The number of Leaf switches is obtained by calculating the ratio of the total number of downlink ports of the Leaf switches to the number of downlink ports of a single Leaf switch; the number of downlink ports of a single Leaf switch is determined based on the total number of GPUs. The number of uplink ports of a single Leaf switch is determined to be the number of Spine switches; based on the symmetrical design of uplink and downlink ports, the number of uplink ports of a single Leaf switch is equal to the number of downlink ports of a single Leaf switch; The network card speed is determined to be the port speed of the Leaf switch and the Spine switch.

[0011] The beneficial effects of adopting the above-mentioned further scheme are as follows: The topology design agent calculates the number of server nodes, the total number of downlink ports on Leaf switches, the total number of uplink ports, the number of Leaf switches, the number of Spine switches, and port speeds sequentially based on key input parameters such as the total number of GPUs, the number of GPUs per node, the number of network interface cards (NICs), NIC speeds, and oversold ratio. The entire process is logically rigorous and computationally transparent. This scheme ensures the accuracy and consistency of topology parameter calculations, avoiding omissions and deviations that may occur during manual calculations.

[0012] Furthermore, the device list generation agent is specifically used for: The number of GPU servers is determined based on the number of server nodes. The specifications and quantity of the Leaf switches and the Spine switches are determined based on the number of Leaf switches and the number of Spine switches. The specifications of the optical module are determined based on the network card speed; The number of first links between the GPU server and the Leaf switch is calculated based on the number of server nodes and the number of network cards in a single server node, and the number of first optical modules between the GPU server and the Leaf switch is calculated based on the number of first links. The number of second links between the Leaf switches and the Spine switches is calculated based on the number of Leaf switches and the number of uplink ports of a single Leaf switch, and the number of second optical modules between the Leaf switches and the Spine switches is calculated based on the number of second links. Add the number of the first optical modules to the number of the second optical modules to get the total number of optical modules; The number of high-speed cables is determined based on the number of server nodes. A bill of materials is generated that includes the specifications and quantities of the GPU server, the Leaf switch, the Spine switch, the optical module, and the high-speed cable.

[0013] The beneficial effects of adopting the above-mentioned further solution are as follows: the equipment bill of materials (BOM) generation agent automatically calculates the specifications and quantities of various equipment based on topology parameters, thus achieving automated BOM generation. This agent can not only determine the quantity of core equipment such as GPU servers, Leaf switches, and Spine switches, but also accurately calculate the number of links between servers and Leaf switches, and between Leaf switches and Spine switches, as well as the corresponding number of optical modules. It automatically summarizes the total number of optical modules and high-speed cables, ultimately generating a complete BOM containing all equipment specifications and quantities. The entire process requires no manual intervention, avoiding omissions and errors that may occur during manual statistics, and providing accurate and standardized material data for engineering procurement and construction.

[0014] Furthermore, the engineering verification layer is specifically used for: Verify the consistency between the number of server nodes and the result obtained by dividing the total number of GPUs by the number of GPUs in a single server node; Verify the consistency between the total number of downlink ports of the Leaf switch and the result obtained by multiplying the number of server nodes by the number of network interface cards (NICs) of a single server node; Verify the consistency between the total number of uplink ports of the Leaf switch and the total number of downlink ports of the Leaf switch divided by the oversold ratio; Verify whether the number of Spine switches is divisible by the number of uplink ports of a single Leaf switch; Verify whether the number of uplink ports of a single Leaf switch is equal to the number of downlink ports of a single Leaf switch; Verify that the number of ports on the Spine switch is not less than the number of ports on the Leaf switch; Verify whether the port speeds of the Leaf switch and the Spine switch are consistent with the network card speed.

[0015] The beneficial effects of adopting the above-mentioned further solution are as follows: The engineering verification layer comprehensively verifies the topology parameters through seven verification rules, ensuring the engineering compliance of the network topology scheme. This solution systematically verifies multiple dimensions, including GPU and node count consistency, Leaf downlink port capacity, oversold ratio consistency, full interconnection conditions, port symmetry, Spine port capacity, and port rate uniformity, covering the main problems that may occur in topology design. Through a rigorous verification mechanism, schemes that do not comply with the design rules can be promptly detected and prevented from entering subsequent stages, effectively avoiding engineering rework and resource waste caused by design errors, and ensuring the accuracy and feasibility of the final output scheme.

[0016] Furthermore, the intelligent agent orchestration layer is implemented through any one of the following methods: workflow, microservice architecture, script flow, and rule engine.

[0017] The beneficial effects of adopting the above-mentioned further solutions are as follows: The intelligent agent orchestration layer provides multiple implementation methods, such as workflow, microservice architecture, script flow, and rule engine, which can be flexibly selected according to actual application scenarios. The workflow approach facilitates visual management and debugging, and is suitable for standardized scenarios with fixed processes; the microservice architecture supports independent deployment and elastic scaling, and is suitable for large-scale production environments; the script flow is lightweight and flexible, suitable for prototype verification and small-scale deployment; the rule engine can automatically match execution logic based on input conditions, and is suitable for dynamic decision-making scenarios. The flexible selection of multiple implementation methods enables this system to adapt to the network construction needs of intelligent computing centers of different scales and stages, improving the system's versatility and practicality.

[0018] Furthermore, when performing topology calculations, the topology design agent calculates the topology parameters using any one of the following methods: rule-driven mathematical models, graph algorithms, heuristic algorithms, and machine learning algorithms.

[0019] The beneficial effects of adopting the above-mentioned further solutions are as follows: the topology design agent provides multiple computational methods, including rule-driven mathematical models, graph algorithms, heuristic algorithms, and machine learning algorithms, which can be flexibly selected according to the actual application scenario. Rule-driven mathematical models have transparent computational logic and deterministic results, making them suitable for scenarios with a high degree of standardization; graph algorithms abstract the network into a graph structure, making them suitable for complex topology optimization scenarios; heuristic algorithms can find optimal or near-optimal solutions under multiple constraints, making them suitable for multi-objective optimization scenarios; machine learning algorithms make predictions based on historical data, making them suitable for scenarios with abundant historical data and requiring rapid response. The flexible selection of multiple algorithms enables this system to obtain optimal topology parameter calculation results under different accuracy, efficiency, and complexity requirements.

[0020] Furthermore, the engineering verification layer performs engineering verification using any one of the following methods: rule engine, constraint solver, graph structure verification algorithm, and network simulation tool.

[0021] The advantages of adopting the above-mentioned further solutions are as follows: The engineering verification layer provides multiple verification methods, including a rule engine, constraint solver, graph structure verification algorithm, and network simulation tool, which can be flexibly selected according to the actual application scenario. The rule engine verifies the topology parameters item by item through preset verification rules, which is simple and efficient; the constraint solver can handle complex constraints and is suitable for multi-variable joint verification scenarios; the graph structure verification algorithm verifies the connectivity and integrity of the topology from a graph theory perspective and is suitable for large-scale network topology verification; the network simulation tool verifies the feasibility of the scheme by simulating network traffic and behavior and is suitable for scenarios with high reliability requirements. The flexible selection of multiple verification methods enables this system to select the most appropriate verification method according to the network design requirements of different scales and complexities, ensuring the engineering compliance of the network topology scheme.

[0022] Furthermore, the user interaction layer receives the computing power requirement information through any one of the following methods: natural language input interface, form-based parameter input interface, structured configuration file interface, and API interface.

[0023] The beneficial effects of adopting the above-mentioned further solutions are as follows: The user interaction layer provides multiple input methods, including a natural language input interface, a form-based parameter input interface, a structured configuration file interface, and an API interface, which can be flexibly selected according to different users' usage habits and scenarios. The natural language input interface allows users to describe their computing power needs in everyday language, lowering the barrier to entry; the form-based parameter input interface ensures the completeness and accuracy of parameters through item-by-item filling, making it suitable for scenarios with a high degree of standardization; the structured configuration file interface supports batch import and automated operation and maintenance, making it suitable for large-scale deployment scenarios; and the API interface facilitates integration with other platforms or automated pipeline calls, making it suitable for inter-system interface scenarios. The flexible selection of multiple input methods enables this system to meet the diverse needs of different user groups, improving the system's usability and adaptability.

[0024] Furthermore, when performing topology drawing, the intelligent agent orchestration layer generates the network topology diagram using any one of the following tools: Mermaid, Graphviz, network design CAD, and Web visualization topology generation tools.

[0025] The beneficial effects of adopting the above-mentioned further solutions are as follows: The topology drawing agent provides multiple drawing methods, including Mermaid, Graphviz, network design CAD, and Web-based visualization topology generation tools, which can be flexibly selected according to actual application scenarios. Mermaid generates topology diagrams using text code, facilitating document embedding and version management; Graphviz supports complex graph layout algorithms, suitable for automatic layout of large-scale topology diagrams; the network design CAD tool provides professional engineering design capabilities, suitable for outputting high-precision engineering drawings; and the Web-based visualization topology generation tool supports interactive browsing and editing, suitable for scheme review and presentation scenarios. The flexible selection of multiple drawing tools enables this system to meet diverse needs for readability, professionalism, and interactivity of topology diagrams in different scenarios.

[0026] Secondly, this invention provides a method for generating a lossless network topology scheme for intelligent computing centers, employing the following technical solution: A method for generating a lossless network topology scheme for an intelligent computing center includes: The computing power demand information is received by the user through the user interaction layer; The intelligent agent orchestration layer drives multiple intelligent agents to sequentially execute requirements parsing, topology calculation, topology drawing, and bill of materials generation; wherein, when performing the topology calculation, the topology design rules stored in the rule knowledge layer are invoked; when performing the bill of materials generation, the bill of materials generation rules stored in the rule knowledge layer are invoked. The topology parameters output by the agent orchestration layer are verified through an engineering verification layer. After successful verification, the network topology scheme is output from the intelligent agent orchestration layer to the user interaction layer for display.

[0027] Thirdly, the present invention provides an electronic device that adopts the following technical solution: An electronic device includes a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the method for generating a lossless network topology scheme for a smart computing center as described in the second aspect.

[0028] Fourthly, the present invention provides a computer-readable storage medium, which adopts the following technical solution: A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for generating a lossless network topology scheme for a smart computing center as described in the second aspect.

[0029] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and will become apparent from the description or may be learned by practice of the invention. Attached Figure Description

[0030] Figure 1 A schematic diagram of the layered architecture of the lossless network topology system for intelligent computing centers provided by the present invention; Figure 2 This is a schematic diagram of the intelligent agent orchestration workflow provided by the present invention; Figure 3 This invention provides a network topology diagram for a small-scale Spine-Leaf two-layer architecture. Figure 4 This invention provides a network topology diagram of a large-scale Spine-Leaf two-layer architecture. Figure 5 A flowchart illustrating the method for generating a lossless network topology scheme for an intelligent computing center provided by the present invention; Figure 6 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0031] The principles and features of the present invention are described below. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0032] Please refer to Figure 1 , Figure 1 This is a schematic diagram of the layered architecture of the lossless network topology generation system for intelligent computing centers provided by this invention. (See diagram below.) Figure 1 As shown, the system may include a user interaction layer 1, an agent orchestration layer 2, a rule knowledge layer 3, and an engineering verification layer 4. The functions and implementation methods of each layer are described in detail below.

[0033] User interaction layer 1 is used to receive computing power demand information input by users and display network topology schemes.

[0034] Optionally, the user interaction layer 1 receives computing power demand information through any one of the following methods: natural language input interface, form-based parameter input interface, structured configuration file interface, and API interface.

[0035] For the natural language input interface, users describe their computing power needs in natural language through text boxes or dialog boxes, for example: I need a training cluster with 1024 GPUs, each server is configured with 8 GPUs and 2 200G network cards, and the oversale ratio is 1:1.

[0036] For form-based parameter input interfaces, users fill in computing power requirement parameters item by item through a web form, such as selecting the total number of GPUs, the number of GPUs per node, and the network card speed from drop-down boxes. The system directly obtains the structured parameters. This approach is suitable for scenarios with clearly defined parameters and a high degree of standardization.

[0037] For the structured configuration file interface, users submit their computing power requirements by uploading configuration files in YAML or JSON format, for example: { "N_gpu": 1024, / / Total number of GPUs "G_per_node": 8, / / Number of GPUs per node "NIC_per_node": 2, / / Number of network cards per node "nic_rate_gbps": 200, / / Network interface card speed (Gbps) "Oversubscription":"1:1" / / Oversubscription ratio } Suitable for batch import or automated operation and maintenance scenarios.

[0038] For API interfaces, third-party systems submit computing power requirements in the form of structured data through RESTful APIs or gRPC interfaces, which are suitable for integration with other platforms or automated pipeline calls.

[0039] The intelligent agent orchestration layer 2, which is connected to the user interaction layer 1, is used to drive multiple intelligent agents to perform requirement analysis, topology calculation, topology drawing and bill of materials generation in sequence.

[0040] Alternatively, the agent orchestration layer 2 can be implemented through any of the following methods: workflow, microservice architecture, script flow, and rule engine.

[0041] For workflow-based approaches, visual workflow engines like Dify Workflow are used to orchestrate agents into sequentially executing nodes. Each node completes a specific task and then passes the result to the next node. For example, after the requirements analysis agent parses the user input into a structured requirements model, it automatically triggers the topology design agent to begin calculations. This approach is easy to manage and debug visually.

[0042] For microservice architecture, each intelligent agent is encapsulated as an independent microservice, and scheduling and communication are handled by a service orchestration engine (such as Kubernetes or Apache Airflow). Each microservice can be deployed and scaled up independently, making it suitable for large-scale production environments.

[0043] For the script-based approach, Python or Shell scripts are written to call the function calls of each agent sequentially. For example, the main control script calls functions such as parse_demand(), compute_topology(), draw_topo(), and generate_bom() in sequence. This approach is lightweight and flexible, suitable for prototype verification and small-scale deployment.

[0044] For rule-based engines, the triggering conditions and execution logic of each agent are defined as rules. The rule engine (such as Drools or EasyRules) automatically matches and executes the corresponding processing logic based on the input conditions. For example, when a structured requirement model is received, the rule engine automatically triggers the topology calculation rule.

[0045] In one embodiment, the agent orchestration layer 2 is specifically used to follow, as shown in the example below. Figure 2 The sequence of agents shown drives the following agents to perform corresponding operations.

[0046] 1) Requirement Analysis Intelligent Agent The demand parsing agent is used to parse computing power demand information into a structured demand model.

[0047] Specifically, the demand parsing agent is based on a large language model, which is guided by preset system prompts to parse the demand according to specified formats and requirements. The system prompts are as follows: You are a computing power demand parser for an intelligent computing center network architecture. Your task is to extract and structure the key information required for the intelligent computing center network design from the user's natural language description. You must: output only JSON format data without any additional explanatory text; if the user does not explicitly provide certain information, complete it reasonably according to common industry default values; all fields must exist and are not allowed to be missing; units should be uniformly Gbps, units, and quantities. Please parse the following user-inputted natural language computing power demand information {{user_input}} into a standard JSON demand description.

[0048] For example, the user's input of natural language computing power requirements is: I need a training cluster with 1024 GPUs, each server configured with 8 GPUs and 2 200G network cards, an oversale ratio of 1:1, a network architecture using Spine-Leaf, a preference for general-purpose vendors, and a medium budget level.

[0049] The demand parsing agent extracts the key information required for the intelligent computing center network design from the natural language computing power demand information input by the user, according to the JSON field definitions shown in Table 1.

[0050] Table 1 JSON Field Definitions The following is an example of the output structured requirements model: { "N_gpu": 1024, "G_per_node": 8, "N_node": 128, "nic_rate_gbps": 200, "NIC_per_node": 2, "network_arch":"spine-leaf", "Oversubscription":"1:1", "vendor_preference":"General", "budget_level":"medium" } At this point, the demand parsing agent has completed the process of converting the user's input natural language computing power demand information into a structured demand model.

[0051] 2) Topology Design Intelligent Agent The topology design agent is used to calculate the topology parameters of the lossless network of the intelligent computing center based on the structured requirements model and topology design rules.

[0052] Specifically, the topology design agent is based on a large language model, which is guided by preset system prompts to perform topology calculations according to specified formats and requirements. The system prompts are as follows: You are a network architecture design expert for an intelligent computing center. Based on the input structured requirement model: total number of GPUs, number of GPUs per node, number of server nodes, number of network cards per node, network card speed, and oversold ratio, design a Spine-Leaf architecture network and output structured JSON topology parameters. Design principles: 1) Each server node uses dual uplinks / multiple uplinks; 2) The number of uplink and downlink ports on the Leaf should be as equal as possible; 3) Full interconnection between the Spine and Leaf nodes; 4) Uniform port speeds.

[0053] The topology design agent receives the structured requirement model output by the requirement parsing agent as input. Following the design principles mentioned above, it calculates the number of server nodes, the total number of downlink ports of Leaf switches, the total number of uplink ports of Leaf switches, the number of Leaf switches, the number of Spine switches, and the port speeds in sequence, and outputs the topology parameters in structured JSON format.

[0054] The following is an example of the output topology parameters: { "architecture":"spine-leaf", "gpu_server": { "count": 128, "port_rate_gbps": 400, "ports": 2 }, "leaf_switch": { "count": 32, "port_rate_gbps": 400, "downlink_ports": 16, "uplink_ports": 16 }, "spine_switch": { "count": 16, "port_rate_gbps": 400, "ports": 32 } } Here, architecture represents the network architecture type; gpu_server represents the GPU server configuration, including the following fields: number of server nodes (count), port rate (port_rate_gbps), and number of ports (ports); leaf_switch represents the Leaf switch configuration, including the following fields: number of Leaf switches (count), port rate (port_rate_gbps), number of downlink ports (downlink_ports), and number of uplink ports (uplink_ports); and spine_switch represents the Spine switch configuration, including the following fields: number of Spine switches (count), port rate (port_rate_gbps), and number of ports (ports).

[0055] How the topology design agent calculates topology parameters using topology design rules is explained in detail in the following section on rule knowledge layer 3.

[0056] Furthermore, when performing topology calculations, the topology design agent is not limited to the rule-driven mathematical model mentioned above, but can also calculate topology parameters through any of the graph algorithms, heuristic algorithms, and machine learning algorithms.

[0057] For rule-driven mathematical models, the topology design agent calculates multiple topology parameters sequentially according to predefined topology design rules based on a structured requirement model. This approach offers transparent computational logic and deterministic results, making it suitable for scenarios with high levels of standardization.

[0058] For graph algorithms, the lossless network of the intelligent computing center is abstracted as a graph structure, where server nodes, Leaf switches, and Spine switches are the vertices of the graph, and network connections are the edges. Algorithms such as graph traversal and graph partitioning are used to automatically generate a topology that satisfies the constraints. This approach is suitable for complex topology optimization scenarios.

[0059] For heuristic algorithms, simulated annealing and genetic algorithms are employed as heuristic search methods to find the optimal or near-optimal combination of topology parameters under constraints such as port capacity and oversold ratio. This approach is suitable for multi-objective optimization scenarios.

[0060] For machine learning algorithms, regression or classification models are trained based on historical network design schemes, and topology parameters are directly predicted after inputting computing power requirements. This approach is suitable for scenarios with abundant historical data and a need for rapid response.

[0061] The topology design agent can flexibly choose any of the above methods to complete the calculation of topology parameters according to the actual application scenario, so as to meet the accuracy, efficiency and complexity requirements of different scenarios.

[0062] 3) Topology drawing agent The topology drawing agent is used to draw the network topology diagram of the lossless network of the intelligent computing center based on topology parameters. The topology drawing agent is only responsible for drawing and does not participate in topology design.

[0063] Specifically, the topology drawing agent is based on a large language model. Preset system prompts guide the model to generate topology graph code according to specified formats and requirements. The system prompts are as follows: You are a network topology drawing assistant. Please generate Mermaid graph TD format topology graph code based on the following structured JSON format topology parameters. Drawing requirements: 1) Display Spine, Leaf, and Server in three layers; 2) Only draw a small number of GPU servers for illustration, using ellipses for the rest; 3) Use concise naming (Spine-1, Leaf-1, Server-1, etc.); 4) Only output Mermaid code, without explanation.

[0064] The topology drawing agent receives the topology parameters output by the topology design agent as input, and uses the Mermaid tool to generate network topology diagram code in Mermaid format according to the above requirements.

[0065] The following is an example of the code for the output Mermaid format network topology diagram: graph TD subgraph Spine_Layer["Spine_Layer"] S1["Spine-1"] S2["Spine-2"] end subgraph Leaf_Layer["Leaf_Layer"] L1["Leaf-1"] L2["Leaf-2"] end subgraph GPU_Server_Layer["GPU_Server_Layer"] G1["GPU-Server-1"] G2["GPU-Server-2"] end S1 -->L1 S1 -->L2 S2 -->L1 S2 -->L2 L1 -->G1 L1 -->G2 L2 -->G1 L2 -->G2 In this context, graph TD represents the top-down topology direction; subgraph is used to define grouping regions, Spine_Layer represents the Spine layer region, Leaf_Layer represents the Leaf layer region, GPU_Server_Layer represents the GPU server layer region; S1 and S2 represent two Spine switches, L1 and L2 represent two Leaf switches, and G1 and G2 represent two GPU servers; the arrow "-->" indicates the network connection relationship between nodes.

[0066] The Mermaid format network topology diagram code based on the above example is shown below. Figure 3 The network topology diagram shown is presented in a three-layer structure, from top to bottom: the Spine layer, the Leaf layer, and the GPU server layer. Spine switches are fully interconnected with Leaf switches, and Leaf switches are fully interconnected with GPU servers.

[0067] In addition to the Mermaid tool mentioned above, the topology drawing agent can also use any of the following tools to generate network topology diagrams: Graphviz, network design CAD tools, and web-based visualization topology generation tools.

[0068] At this point, the topology drawing agent has completed the process of drawing a network topology graph based on topology parameters.

[0069] 4) Intelligent agent for generating equipment list The equipment list generation agent is used to generate the bill of materials for the lossless network of the intelligent computing center based on topology parameters and bill of materials generation rules.

[0070] Specifically, the equipment list generation agent is based on a large language model. It guides the large language model to generate an equipment bill of materials according to specified formats and requirements through preset system prompts. The system prompt is as follows: You are a network equipment planning expert for the intelligent computing center. Please generate an equipment bill of materials based on the topology parameters. Generation rules: 1) Servers, Leaf, and Spine are listed separately; 2) Number of optical modules = Total number of ports; 3) Number of high-speed cables = Number of optical modules / 2; 4) Only output the equipment bill of materials in JSON format.

[0071] The equipment list generation agent receives the topology parameters output by the topology design agent as input and generates a JSON-formatted equipment bill of materials according to the above rules.

[0072] The output equipment bill of materials example is as follows: { "devices": [ { "name":"GPU Server", "spec":"8xH800 GPU", "quantity": 128 }, { "name":"Leaf Switch", "spec":"32x400G", "quantity": 32 }, { "name":"Spine Switch", "spec":"64x400G", "quantity": 16 }, { "name":"400G Optical Module", "spec":"DR4", "quantity": 1024 }, { "name":"High-Speed ​​Cable", "spec":"30m", "quantity": 512 } ] } The "devices" field lists the devices, including five types: GPU Server, Leaf Switch, Spine Switch, Optical Module, and High-Speed ​​Cable. Each device type contains three fields: name, specification, and quantity.

[0073] The equipment bill of materials (BOM) generation agent calculates the specifications and quantities of GPU servers, Leaf switches, Spine switches, optical modules, and high-speed cables using the BOM generation rules to generate the equipment BOM. For details on the rule knowledge layer 3, please refer to the following section.

[0074] At this point, the equipment bill of materials generation agent has completed the process of generating the bill of materials based on topology parameters.

[0075] 5) Solution Output Intelligent Agent The solution outputs an intelligent agent, which, after the topology parameters have been verified, generates a network topology scheme based on the structured requirements model, the topology parameters, the network topology diagram, and the bill of materials.

[0076] Specifically, the solution output agent is based on a large language model. It guides the model to generate a network topology scheme according to specified formats and requirements through preset system prompts. The system prompts are as follows: You are the network solution output assistant for the intelligent computing center. Please generate a complete network topology scheme based on the following: 1) Solution overview; 2) Network architecture description; 3) Network topology diagram (directly embedding Mermaid format network topology diagram code); 4) Equipment list table. The tone is professional and restrained, without making any sales promises.

[0077] The solution output agent receives the following four inputs: 1) Requirements: The structured requirements model output by the requirements parsing agent; 2) Topology: The topology parameters output by the topology design agent; 3) Topology diagram: The network topology diagram output by the topology drawing agent; 4) Bill of materials: The equipment bill of materials output by the equipment generating agent.

[0078] The solution output agent integrates the above four inputs into a complete network design solution, including a solution overview, network architecture description, network topology diagram and equipment list table, and outputs it in a professional and restrained tone without making any sales promises.

[0079] At this point, the solution output agent has completed the process of generating the network topology scheme, which will be output to the user interaction layer 1 for display.

[0080] The verification process for the above topology parameters is performed by the network topology verification node in the engineering verification layer 4. See the following description of the engineering verification layer 4 for details.

[0081] The rule knowledge layer 3, which is connected to the agent orchestration layer 2, is used to store and provide topology design rules and bill of materials generation rules, so that the agent orchestration layer 2 can call the topology design rules when performing topology calculations and the bill of materials generation rules when performing bill of materials generation.

[0082] Specifically, the knowledge base in rule knowledge layer 3 is expert experience. Hardcoding rules directly into system prompts is inconvenient for maintenance; therefore, a knowledge base is used to store rule-type documents, decoupling rules from agents. The system prompts for agents explicitly state: "Decisions can only be made based on rules in the knowledge base; no rules can be fabricated," ensuring the standardization and consistency of the network topology scheme.

[0083] Rule knowledge layer 3 stores topology design rules and device list generation rules. The following detailed explanations of each rule and its application, using specific examples, illustrate these rules.

[0084] I. Topology Design Rules The topology design rules include Leaf switch port calculation rules and Spine switch port calculation rules.

[0085] (a) Leaf Switch Port Calculation Rules 1) Leaf / Spine calculation approach The Leaf / Spine computing approach involves the following parameters: total number of GPUs, number of GPUs per node, number of server nodes, number of network interface cards (NICs) per node, NIC speed, and oversold ratio. These parameters are derived from the structured requirement model output by the requirement analysis agent.

[0086] The calculation steps are as follows: a) Calculate the total number of GPUs; b) Number of nodes = Total number of GPUs / Number of GPUs per node; c) Node injection bandwidth = Number of NICs × NIC rate; d) Number of Leaf downlink ports = Number of nodes / Number of Leafs; e) Leaf uplink port count = downlink port count / oversubscription ratio; f) Number of Spine ports = Number of Leaf uplink ports.

[0087] The above calculations are based on the following fundamental assumptions: a) the network structure is a two-layer Spine-Leaf structure; b) the nodes are connected to dual network cards and dual ToRs; c) the Leaf-Spine is uniformly interconnected; d) the oversubscription ratio is 1:1 by default; e) the Spine does not host any servers and is a pure aggregation.

[0088] 2) Leaf switch port calculation rules Input variables: total number of GPUs (N_gpu), number of GPUs per node (G_per_node), number of server nodes (N_node=N_gpu / G_per_node), number of network cards per node (NIC_per_node).

[0089] Leaf device selection: a) When the computing power requirement is no more than 1024 GPUs, select a 32×400G Leaf device with 16 uplinks and 16 downlinks; b) When the computing power requirement is more than 1024 GPUs, select a 64×400G Leaf device with 32 uplinks and 32 downlinks.

[0090] Leaf downlink ports (connected to servers): One node occupies two or more Leaf ports to avoid single points of failure. The total number of Leaf switch downlink ports is: Leaf_downlink_ports_total = N_node × NIC_per_node.

[0091] Leaf uplink ports (connected to Spine): Determined by the oversubscription ratio. The total number of Leaf uplink ports is: Leaf_uplink_ports_total = Leaf_downlink_ports_total / Oversubscription. An oversubscription ratio of 1:1 is suitable for LLM training, and 2:1 is suitable for inference / mixed workloads.

[0092] Number of ports per Leaf switch: Leaf switches should adopt a symmetrical design for uplink and downlink ports, with the number of downlink ports per Leaf switch being approximately equal to the number of uplink ports. The number of ports per Leaf switch is: Leaf_ports_per_switch = 2 × max(Leaf_downlink_ports_per_switch, Leaf_uplink_ports_per_switch). Where Leaf_ports_per_switch represents the total number of ports per Leaf switch, Leaf_downlink_ports_per_switch represents the number of downlink ports per Leaf switch, and Leaf_uplink_ports_per_switch represents the number of uplink ports per Leaf switch.

[0093] Leaf Count Calculation: Assume a 32-port or 64-port 400G Leaf device is selected. Taking a 32-port Leaf device as an example, it is configured with 16 downlink ports and 16 uplink ports. The number of Leaf switches is: Leaf_count = ceil(Leaf_downlink_ports_total / 16). Where Leaf_count represents the number of Leaf switches, Leaf_downlink_ports_total represents the total number of downlink ports on the Leaf switches, 16 represents the number of downlink ports on a single 32-port Leaf device, and ceil represents rounding up.

[0094] For example, taking 1024 GPUs as an example, the total number of GPUs is 1024, the number of GPUs per node is 8, then the number of server nodes is 1024 / 8=128; the number of network cards per node is 2, then the total number of downlink ports of Leaf switches is 128×2=256; if a 32-port Leaf device (16 downlinks) is selected, then the number of Leaf switches is 256 / 16=16.

[0095] Based on the aforementioned Leaf switch port calculation rules, when performing topology calculations, the topology design agent obtains the number of server nodes by the ratio of the total number of GPUs to the number of GPUs per node; the total number of Leaf switch downlink ports is obtained by multiplying the number of server nodes by the number of network cards per node; the total number of Leaf switch uplink ports is obtained by the ratio of the total number of Leaf switch downlink ports to the oversubscription ratio; and the number of Leaf switches is obtained by the ratio of the total number of Leaf switch downlink ports to the number of downlink ports per Leaf switch, where the number of downlink ports per Leaf switch is determined based on the total number of GPUs.

[0096] (II) Spine Switch Port Calculation Rules Spine's basic rule: Number of Spine switches = Number of uplink ports per Leaf.

[0097] Spine equipment selection: a) When the maximum number of leaves does not exceed 32, select a 32×400G Spine equipment; b) When the maximum number of leaves does not exceed 64, select a 64×400G Spine equipment.

[0098] The calculation formula is: Spine_count = Leaf_uplink_ports_per_switch. Where Spine_count represents the number of Spine switches, and Leaf_uplink_ports_per_switch represents the number of uplink ports per Leaf switch.

[0099] For example, if each Leaf has 16 uplink ports, then the number of Spine switches is 16.

[0100] It's important to note that, to save costs, the uplink ports of the Leaf switch can be divided into several groups, with each group connected to one Spine switch. This makes the rule that the number of Spine switches is divisible by the number of uplink ports on each Leaf switch.

[0101] Based on the above Spine switch port calculation rules, when performing topology calculations, the topology design agent determines the number of uplink ports of a single Leaf switch to be equal to the number of Spine switches. Based on the symmetrical design of uplink and downlink ports, the number of uplink ports of a single Leaf switch is equal to the number of downlink ports. The network card speed is determined to be equal to the port speed of the Leaf switch and the Spine switch.

[0102] II. Equipment List Generation Rules The equipment list generation rules include calculation rules for server-Leaf optical modules, calculation rules for Leaf-Spine optical modules, and calculation rules for other equipment.

[0103] (I) Server-Leaf Optical Module Calculation Rules 1) Calculation approach for optical modules General principle: Number of optical modules = Number of physical links × Number of optical modules per link. Link types include server-Leaf and Leaf-Spine.

[0104] Optical module model selection: a) When the distance is less than 500m (within the same equipment room), select the 400G SR4 optical module; b) When the distance is less than 2km, select the 400G DR4 optical module.

[0105] 2) Server-Leaf optical module calculation rules (downlink) Basic rules: Each NIC corresponds to one Leaf port, and there is one optical module at each end of a link.

[0106] Calculation formula: Server_Link_Count=N_node×NIC_per_node; Server_Optical_Module_Count=Server_Link_Count×2; Here, Server_Link_Count represents the number of links from the server to the Leaf switch, and Server_Optical_Module_Count represents the number of optical modules on the server side.

[0107] For example, taking 1024 GPUs as an example, the total number of GPUs is 1024, the number of GPUs per node is 8, then the number of server nodes is 1024 / 8=128; the number of network cards per node is 2, then the number of server links is 128×2=256, and the number of server-side optical modules is 256×2=512.

[0108] Based on the above server-Leaf optical module calculation rules, the device list generation agent calculates the number of first links between the GPU server and the Leaf switch according to the number of server nodes and the number of network cards per node, and calculates the number of first optical modules according to the number of first links.

[0109] (II) Leaf-Spine optical module calculation rules (uplink) The physical structure of Leaf-Spine: Each Leaf switch has several uplink ports (uplink_ports_per_leaf), and each uplink port is connected to a different Spine switch. Each link requires one optical module at each end.

[0110] Calculation formula: Leaf_Spine_Link_Count=Leaf_count×uplink_ports_per_leaf; Leaf_Spine_Optical_Module_Count=Leaf_Spine_Link_Count×2; Wherein, Leaf_Spine_Link_Count represents the number of links from the Leaf switch to the Spine switch, and Leaf_Spine_Optical_Module_Count represents the number of optical modules on the Leaf-Spine side.

[0111] For example, taking 1024 GPUs as an example, the number of Leaf switches is 16, the number of uplink ports per Leaf switch is 16, then the number of Leaf-Spine links is 16 × 16 = 256, and the number of Leaf-Spine optical modules is 256 × 2 = 512.

[0112] Based on the Leaf-Spine optical module calculation rules mentioned above, the device inventory generation agent calculates the number of second links between Leaf switches and Spine switches according to the number of Leaf switches and the number of uplink ports on a single Leaf switch, and then calculates the number of second optical modules based on the number of second links.

[0113] Total number of optical modules = number of server optical modules + number of Leaf-Spine optical modules.

[0114] Based on this rule, the device list generating agent adds the number of the first optical modules to the number of the second optical modules to obtain the total number of optical modules.

[0115] (III) Optical Module Verification Rules After the optical module calculation is completed, the following verifications need to be performed: Verification 1: Parity check of the total number of optical modules Verification formula: Total_Optical_Modules%2=0; Verification objective: To ensure that the total number of optical modules is even.

[0116] Verification 2: Leaf-Spine optical module quantity verification Verification formula: Leaf_Spine_Optical_Module_Count = Leaf_count × Leaf_uplink_ports_per_switch × 2; Verification objective: To ensure that the number of Leaf-Spine optical modules equals the number of ports on the Spine side multiplied by 2.

[0117] Verification 3: Verification of the number of optical modules on the server side Verification formula: Server_Optical_Module_Count = N_node × NIC_per_node × 2; Verification objective: To ensure that the number of optical modules on the server side equals the number of nodes multiplied by the number of network cards per node, and then multiplied by 2.

[0118] (iv) Calculation rules for other equipment Other device calculation rules are used to guide the device list generation agent to determine the specifications and quantity of devices other than optical modules.

[0119] The number of GPU servers is determined by the number of server nodes in the topology parameters; that is, the number of GPU servers equals the number of server nodes.

[0120] Based on this rule, the device inventory generation agent determines the number of GPU servers according to the number of server nodes.

[0121] The specifications and quantity of Leaf switches are determined based on the number of Leaf switches and port speeds in the topology parameters. That is, the number of Leaf switches is equal to the number of Leaf switches in the topology parameters, and the port speed is equal to the network card speed.

[0122] Based on this rule, the device list generating agent determines the specifications and quantity of Leaf and Spine switches according to the number of Leaf switches and Spine switches.

[0123] The specifications and number of Spine switches are determined based on the number of Spine switches and port speeds in the topology parameters. That is, the number of Spine switches is equal to the number of Spine switches in the topology parameters, and the port speed is equal to the network card speed.

[0124] The number of high-speed cables is determined based on the number of server nodes and the number of network interface cards (NICs) per node; that is, the number of high-speed cables = the number of server nodes × the number of NICs per node. Based on this rule, the device inventory generation agent determines the number of high-speed cables according to the number of server nodes.

[0125] Based on the aforementioned calculation rules for server-Leaf optical modules, Leaf-Spine optical modules, and other equipment calculation rules, the equipment list generation agent determines the specifications of the optical modules according to the network card speed, and compiles and generates a bill of materials containing the specifications and quantities of GPU servers, Leaf switches, Spine switches, optical modules, and high-speed cables.

[0126] Engineering verification layer 4, which is connected to agent orchestration layer 2 and user interaction layer 1, is used to perform engineering verification on the topology parameters output by agent orchestration layer 2, and trigger agent orchestration layer 2 to output network networking scheme to user interaction layer 1 after the verification is passed.

[0127] Specifically, the engineering verification layer 4 includes a network topology verification node, which verifies whether the topology parameters conform to the topology design rules. If the topology parameters pass the verification, the intelligent agent orchestration layer 2 is triggered to output the network topology scheme to the user interaction layer 1 for display; if the topology parameters fail the verification, the scheme generation process is terminated and no network topology scheme is output.

[0128] Optionally, the engineering verification layer 4 performs engineering verification using any one of the following methods: rule engine, constraint solver, graph structure verification algorithm, and network simulation tool.

[0129] The following example uses the rule engine. The network topology verification node in the engineering verification layer 4 verifies the topology parameters according to the following verification rules through the rule engine.

[0130] Verification Rule 1: Consistency Verification of GPU and Node Count Verification formula: N_node × G_per_node == N_gpu; The == operator is used to compare the equality of two values.

[0131] Verification objective: To verify the consistency between the number of server nodes and the total number of GPUs divided by the number of GPUs per server node. If the result is inconsistent, the verification is deemed a failure.

[0132] Verification Rule 2: Leaf Downlink Port Capacity Verification Verification formula: Leaf_count × Leaf_downlink_ports_per_switch ≥ N_node × NIC_per_node; Where Leaf_count represents the number of Leaf switches, Leaf_downlink_ports_per_switch represents the number of downlink ports per Leaf switch, N_node represents the number of server nodes, and NIC_per_node represents the number of network interface cards per node.

[0133] Purpose of verification: To verify the consistency between the total number of downlink ports on the Leaf switch and the result obtained by multiplying the number of server nodes by the number of network interface cards (NICs) on a single server node. If the result is inconsistent, it indicates that there are not enough ports.

[0134] Verification Rule 3: Divisibility check between Leaf uplink port and Spine number (full interconnection) Verification formula: Leaf_uplink_ports_per_switch%Spine_count==0; Verification purpose: To verify whether the number of Spine switches is divisible by the number of uplink ports of a single Leaf switch, ensuring full interconnection between Leaf and Spine switches. If this condition is not met, full interconnection cannot be achieved.

[0135] Verification Rule 4: Spine Port Capacity Verification Verification formula: Leaf_count ≤ Spine_ports_per_switch; Verification purpose: To verify that the number of ports on the Spine switch is not less than the number of Leaf switches, ensuring that the Spine switch has enough ports to connect all Leaf switches. If this is not the case, it indicates that the Spine has insufficient ports.

[0136] Verification Rule 5: Leaf Port Symmetry Verification Verification formula: Leaf_downlink_ports_per_switch==Leaf_uplink_ports_per_switch; Verification Purpose: To verify whether the number of uplink ports and downlink ports of a single Leaf switch are equal to meet the standard ToR design. If they do not match, it is a non-standard design and a warning will be issued.

[0137] Verification Rule 6: Overselling Ratio Consistency Verification Verification formula: Leaf_downlink_ports_per_switch / Leaf_uplink_ports_per_switch==Oversubscription; Verification objective: To verify the consistency between the total number of uplink ports on the Leaf switch and the total number of downlink ports divided by the oversold ratio. If they do not match, it indicates an inconsistency in the bandwidth model, and a warning will be issued.

[0138] Verification Rule 7: Size Limit Protection Verification Verification formula: Leaf_count≤64 AND Spine_count≤64; Purpose of validation: To prevent large language models from creating illusions that exceed the scale supported by the current model. If the validation fails, return `unsupported_scale=true`, indicating validation failure.

[0139] Verification Rule 8: Port Rate Consistency Verification Verification formula: Leaf_port_rate == nic_rate AND Spine_port_rate == nic_rate_gbps; Verification objective: To verify whether the port speeds and network interface card (NIC) speeds of the Leaf and Spine switches are consistent. If they are not consistent, the port speeds are mismatched, and the verification fails.

[0140] After performing the verification according to the above verification rules, the network topology verification node returns the verification result, including: valid: A boolean value indicating whether the validation passed; errors: An error list that records all validation rules that failed; Warnings: A list of warnings that records all non-standard design issues that do not reach the level of failure.

[0141] If the errors list is empty, the verification passes and the intelligent agent orchestration layer 2 is triggered to output the network topology scheme to the user interaction layer 1 for display; if the errors list is not empty, the verification fails and the process is terminated.

[0142] The lossless network topology generation system for intelligent computing centers provided by this invention achieves automated design of lossless network topology schemes for intelligent computing centers through a layered architecture design consisting of a user interaction layer, an intelligent agent orchestration layer, a rule knowledge layer, and an engineering verification layer. Users only need to input computing power requirements, and the system can automatically complete requirements parsing, topology calculation, topology drawing, and bill of materials generation. The engineering verification layer automatically verifies the topology parameters to ensure the engineering compliance of the generated scheme. Compared to traditional network planning methods that rely on manual experience, this invention significantly improves the design efficiency and standardization of network topology schemes, reduces the risks of human design, and can meet the needs of rapid construction and large-scale replication of intelligent computing centers.

[0143] The following detailed description of the lossless network topology generation system for intelligent computing centers provided by the present invention is based on a specific embodiment.

[0144] This example uses the network requirements of a university's intelligent computing center. The customer needs 8 servers, each containing 8 NVIDIA H100 graphics cards, for a total of 64 graphics cards. Each server is configured with 4 400G interfaces, and the network adopts a Spine-Leaf architecture with an oversold ratio of 1:1 and no convergence.

[0145] The user inputs their computing power requirements through the natural language input interface of the user interaction layer: I need 8 servers, each containing 8 NVIDIA H100 graphics cards, for a total of 64 graphics cards. Each server has 4 400G interfaces, and the network topology should use a Spine-Leaf architecture with 1:1 non-convergence. Please help design a network topology scheme.

[0146] The output network topology scheme is as follows: Lossless Network Deployment Solution for Intelligent Computing Center 1) Overview of the Plan This solution aims to build a high-performance, non-blocking dedicated network for an intelligent computing cluster comprising 64 high-performance GPUs (H100). Strictly adhering to user requirements, the solution employs a Spine-Leaf CLOS architecture, achieving a 1:1 convergence-free design to ensure 400Gbps end-to-end bidirectional bandwidth between any GPU servers. This meets the low-latency, high-throughput network requirements of large-scale distributed AI training and high-performance computing scenarios. Equipment selection is based on the user-specified NVIDIA brand, with a mid-range budget, configuring the network while meeting performance targets.

[0147] 2) Network Architecture Description Architecture type: Two-layer Spine-Leaf CLOS architecture. This architecture has advantages such as strong horizontal scalability, high path redundancy, and predictable latency, and is the standard network design for high-performance computing and AI clusters.

[0148] Non-blocking design: By configuring 4 Leaf switches and 4 Spine switches and using full-mesh interconnection, a strict 1:1 non-convergence ratio is achieved. The total bandwidth (1.6Tbps) of the four 400G downlinks of each server node is perfectly matched with the total bandwidth of the uplinks connected to all Spine switches, eliminating network-level performance bottlenecks.

[0149] Traffic model: All inter-server communication across Leaf (East-West traffic) is forwarded through the Spine layer. Leaf switches act as gateways for servers, and Spine switches act as the core switching layer, together ensuring equivalent and optimal multipath forwarding capabilities between any two points.

[0150] Scalability: The current architecture provides a clear path for expansion. Uplink bandwidth and more Leaf switches can be connected by adding Spine switches; more compute nodes can be connected by adding Leaf switches, while maintaining the symmetry and non-blocking characteristics of the architecture.

[0151] 3) Network topology diagram The generated network topology diagram is as follows Figure 4 The diagram shows the logical network topology and connections of this network topology scheme.

[0152] 4) Equipment List Table 2 below lists the main hardware devices, specifications, and quantities required to implement this network topology scheme.

[0153] Table 2 Equipment List This embodiment fully verifies the significant effect of the present invention in improving the design efficiency of intelligent computing center network topology schemes and reducing human design risks. It can be widely applied to scenarios such as operator intelligent computing center planning, research institution AI cluster construction, and government and enterprise computing power infrastructure scheme design.

[0154] Please refer to Figure 5 , Figure 5 This is a flowchart illustrating the method for generating a lossless network topology scheme for intelligent computing centers provided by the present invention. Figure 5 As shown, the method may include the following steps: S501, Receive computing power demand information input by the user through the user interaction layer; S502, Drive multiple agents through the agent orchestration layer to sequentially execute requirement parsing, topology calculation, topology drawing and bill of materials generation; Specifically, when performing topology calculations, the topology design rules stored in the rule knowledge layer are called; when performing bill of materials generation, the bill of materials generation rules stored in the rule knowledge layer are called. S503. Perform engineering verification on the topology parameters output by the agent orchestration layer through the engineering verification layer; S504. After the verification is passed, the network topology scheme is output to the user interaction layer through the intelligent agent orchestration layer for display.

[0155] In some embodiments, the intelligent computing center lossless network topology generation system of the present invention can be implemented in a combination of hardware and software. As an example, the intelligent computing center lossless network topology generation system of the present invention can be a processor in the form of a hardware decoding processor, which is programmed to execute the intelligent computing center lossless network topology generation method of the present invention. For example, the processor in the form of a hardware decoding processor can be one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components.

[0156] The modules described in the embodiments of this invention can be implemented in software or hardware. The names of the modules are not, in some cases, limiting the scope of the module itself.

[0157] An electronic device according to an embodiment of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-described method for generating a lossless network topology scheme for a smart computing center. That is, an electronic device according to an embodiment of the present invention may include, but is not limited to: a processor and a memory; the memory is used to store the computer program; the processor is used to execute the method for generating a lossless network topology scheme for a smart computing center as shown in the present invention by calling the computer program.

[0158] In one alternative embodiment, an electronic device is provided, such as Figure 6 As shown, Figure 6 The illustrated electronic device 4000 includes a processor 4001 and a memory 4003. The processor 4001 and the memory 4003 are connected, for example, via a bus 4002. Optionally, the electronic device 4000 may further include a transceiver 4004, which can be used for data interaction between the electronic device and other electronic devices, such as sending and / or receiving data. It should be noted that in practical applications, the transceiver 4004 is not limited to one type, and the structure of the electronic device 4000 does not constitute a limitation on the embodiments of the present invention.

[0159] Processor 4001 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this invention. Processor 4001 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0160] Bus 4002 may include a path for transmitting information between the aforementioned components. Bus 4002 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. Bus 4002 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 6 The bus 4002 is represented by only one thick line, but this does not mean that there is only one bus or one type of bus.

[0161] The memory 4003 may be ROM (Read Only Memory) or other types of static storage devices capable of storing static information and instructions, RAM (Random Access Memory) or other types of dynamic storage devices capable of storing information and instructions, or EEPROM (Electrically Erasable Programmable Read Only Memory), CD-ROM (Compact Disc Read Only Memory) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto.

[0162] The memory 4003 stores application code (computer program) for executing the present invention, and its execution is controlled by the processor 4001. The processor 4001 executes the application code stored in the memory 4003 to implement the content shown in the foregoing method embodiments.

[0163] Among them, electronic devices can also be terminal devices, which can be any device that can install applications, including at least one of smartphones, tablets, laptops, desktop computers, smart speakers, smartwatches, smart TVs, and smart in-vehicle devices.

[0164] It should be noted that, Figure 6 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0165] An embodiment of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described method for generating a lossless network topology scheme for a smart computing center.

[0166] Alternatively, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), magnetic tape, a floppy disk, and an optical data storage device, etc.

[0167] In an exemplary embodiment, a computer program product or computer program is also provided, which includes computer instructions stored in a computer-readable storage medium. The processor of an electronic device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the electronic device to perform the aforementioned method for generating a lossless network topology scheme for a smart computing center.

[0168] Computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0169] It should be understood that the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0170] The computer-readable storage medium provided in this invention can be, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EEPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0171] The aforementioned computer-readable storage medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the method shown in the above embodiments.

[0172] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this invention is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this invention.

[0173] It should be noted that the terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and represent a limitation on a specific order or sequence. Where appropriate, the order of use for similar objects can be interchanged so that the embodiments of this application described herein can be implemented in an order other than that shown or described.

[0174] Those skilled in the art will recognize that this invention can be implemented as a system, method, or computer program product. Therefore, this invention can be specifically implemented in the following forms: it can be entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software, generally referred to herein as a "circuit," "module," or "system." Furthermore, in some embodiments, this invention can also be implemented as a computer program product contained in one or more computer-readable media, which includes computer-readable program code.

[0175] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A system for generating lossless network topology schemes for intelligent computing centers, characterized in that, include: The user interaction layer is used to receive computing power demand information input by users and display network topology schemes. The intelligent agent orchestration layer, which is connected to the user interaction layer, is used to drive multiple intelligent agents to sequentially execute requirement parsing, topology calculation, topology drawing and bill of materials generation. The rule knowledge layer, which is connected to the agent orchestration layer, is used to store and provide topology design rules and bill of materials generation rules, so that the agent orchestration layer can call the topology design rules when performing the topology calculation and the bill of materials generation rules when performing the bill of materials generation. The engineering verification layer, which is connected to the agent orchestration layer and the user interaction layer, is used to perform engineering verification on the topology parameters output by the agent orchestration layer, and triggers the agent orchestration layer to output the network topology scheme to the user interaction layer after the verification is passed.

2. The intelligent computing center lossless network topology generation system according to claim 1, characterized in that, The agent orchestration layer is specifically used to drive the following multiple agents to perform corresponding operations according to the agent orchestration order: A demand parsing intelligent agent is used to parse the computing power demand information into a structured demand model; A topology design agent is used to calculate the topology parameters of the lossless network of the intelligent computing center based on the structured requirements model and the topology design rules. A topology drawing agent is used to draw a network topology diagram of the lossless network of the intelligent computing center based on the topology parameters. An intelligent agent for generating equipment lists is used to generate a bill of materials for the lossless network of the intelligent computing center based on the topology parameters and the bill of materials generation rules. The solution outputs an intelligent agent, which, after the topology parameters have passed verification, generates the network topology scheme based on the structured requirements model, the topology parameters, the network topology diagram, and the bill of materials.

3. The intelligent computing center lossless network topology generation system according to claim 2, characterized in that, The structured requirements model includes the total number of GPUs, the number of GPUs and network cards per server node, network card speed, and oversold ratio; the topology parameters include the number of server nodes, the number of Leaf switches, the number of uplink and downlink ports and port speeds, and the number of Spine switches and port speeds. The topology design agent is specifically used for: The number of server nodes is obtained by calculating the ratio between the total number of GPUs and the number of GPUs in a single server node; The total number of downlink ports of the Leaf switch is obtained by multiplying the number of server nodes by the number of network interface cards (NICs) of a single server node. Calculate the ratio between the total number of downlink ports of the Leaf switch and the oversold ratio to obtain the total number of uplink ports of the Leaf switch; The number of Leaf switches is obtained by calculating the ratio of the total number of downlink ports of the Leaf switches to the number of downlink ports of a single Leaf switch; the number of downlink ports of a single Leaf switch is determined based on the total number of GPUs. The number of uplink ports of a single Leaf switch is determined as the number of Spine switches; Based on the symmetrical design of uplink and downlink ports, the number of uplink ports of a single Leaf switch is equal to the number of downlink ports of a single Leaf switch; The network card speed is determined to be the port speed of the Leaf switch and the Spine switch.

4. The intelligent computing center lossless network topology generation system according to claim 3, characterized in that, The device list generation agent is specifically used for: The number of GPU servers is determined based on the number of server nodes. The specifications and quantity of the Leaf switches and the Spine switches are determined based on the number of Leaf switches and the number of Spine switches. The specifications of the optical module are determined based on the network card speed; The number of first links between the GPU server and the Leaf switch is calculated based on the number of server nodes and the number of network cards in a single server node, and the number of first optical modules between the GPU server and the Leaf switch is calculated based on the number of first links. The number of second links between the Leaf switches and the Spine switches is calculated based on the number of Leaf switches and the number of uplink ports of a single Leaf switch, and the number of second optical modules between the Leaf switches and the Spine switches is calculated based on the number of second links. Add the number of the first optical modules to the number of the second optical modules to get the total number of optical modules; The number of high-speed cables is determined based on the number of server nodes. A bill of materials is generated that includes the specifications and quantities of the GPU server, the Leaf switch, the Spine switch, the optical module, and the high-speed cable.

5. The intelligent computing center lossless network topology generation system according to claim 3, characterized in that, The engineering verification layer is specifically used for: Verify the consistency between the number of server nodes and the result obtained by dividing the total number of GPUs by the number of GPUs in a single server node; Verify the consistency between the total number of downlink ports of the Leaf switch and the result obtained by multiplying the number of server nodes by the number of network interface cards (NICs) of a single server node; Verify the consistency between the total number of uplink ports of the Leaf switch and the total number of downlink ports of the Leaf switch divided by the oversold ratio; Verify whether the number of Spine switches is divisible by the number of uplink ports of a single Leaf switch; Verify whether the number of uplink ports of a single Leaf switch is equal to the number of downlink ports of a single Leaf switch; Verify that the number of ports on the Spine switch is not less than the number of ports on the Leaf switch; Verify whether the port speeds of the Leaf switch and the Spine switch are consistent with the network card speed.

6. The intelligent computing center lossless network topology generation system according to claim 1, characterized in that, The intelligent agent orchestration layer is implemented through any one of the following methods: workflow, microservice architecture, script flow, and rule engine.

7. The intelligent computing center lossless network topology generation system according to claim 2, characterized in that, When performing topology calculations, the topology design agent calculates the topology parameters using any one of the following methods: rule-driven mathematical models, graph algorithms, heuristic algorithms, and machine learning algorithms.

8. The intelligent computing center lossless network topology generation system according to claim 1, characterized in that, The engineering verification layer performs engineering verification using any one of the following methods: rule engine, constraint solver, graph structure verification algorithm, and network simulation tool.

9. The intelligent computing center lossless network topology generation system according to claim 1, characterized in that, The user interaction layer receives the computing power requirement information through any one of the following methods: natural language input interface, form-based parameter input interface, structured configuration file interface, and API interface.

10. The intelligent computing center lossless network topology generation system according to claim 1, characterized in that, When performing topology drawing, the intelligent agent orchestration layer generates the network topology diagram using any one of the following tools: Mermaid, Graphviz, network design CAD, and Web visualization topology generation tools.

11. A method for generating a lossless network topology scheme for an intelligent computing center, characterized in that, include: The computing power demand information is received by the user through the user interaction layer; The intelligent agent orchestration layer drives multiple intelligent agents to sequentially execute requirements parsing, topology calculation, topology drawing, and bill of materials generation; wherein, when performing the topology calculation, the topology design rules stored in the rule knowledge layer are invoked; when performing the bill of materials generation, the bill of materials generation rules stored in the rule knowledge layer are invoked. The topology parameters output by the agent orchestration layer are verified through an engineering verification layer. After successful verification, the network topology scheme is output from the intelligent agent orchestration layer to the user interaction layer for display.

12. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the method for generating a lossless network topology scheme for a smart computing center as described in claim 11.

13. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the method for generating a lossless network topology scheme for a smart computing center as described in claim 11.

Images