Adaptive network topology detection method based on AI computing power data center

Abstract

The invention discloses a self-adaptive network topology detection method based on an AI computing power data center, and the method comprises the steps: S1, the starting of data center physical equipment and the communication preparation of a cable, and guaranteeing the connection of an equipment physical link of an AI computing power data center network; step S2, deploying an agent program of an LTDP (Link Topology Discovery Protocol) protocol, which supports protocol extension based on an LLDP (Link Layer Discovery Protocol), for switch equipment of the computing power network data center, and deploying an agent program of an LTDP (Link Topology Discovery Protocol) protocol, which supports protocol extension based on the LLDP (Link Layer Discovery Protocol); s3, starting an agent program on the switch equipment to start to collect local LTDP information, wherein the content comprises LLDP information and computing power network characteristic information related to interfaces, flow control and equipment; and S4, deploying the adaptive network topology detection device at a universal server node of a computing power network data center, and starting the adaptive network topology detection device. And after receiving the TDP response message, the topology detection device continues to construct the TDP of the far-end equipment, so that intelligent detection and information collection of network equipment topology in the computing power center are realized.

Description

Technical Field

[0001] This invention relates to the field of data network communication, and in particular to an adaptive network topology detection method based on AI computing power data centers. Background Technology

[0002] In traditional data center Ethernet architectures, networks typically rely on distributed protocols (such as OSPF and BGP) for route convergence and path selection, while operations and maintenance depend on manual planning of VLANs, VXLANs, ACLs, QoS policies, and node address allocation. With the increasing deployment of AI, HPC, and large-scale storage systems, this model has exposed numerous problems. First, network topology scales rapidly; common spine-leaf or CLOS architectures may involve hundreds or thousands of switching nodes and thousands of servers, making traditional manual configuration or script-based automation methods inefficient and inaccurate. Second, uncontrollable paths and a lack of a global perspective have become major pain points. Under ECMP or static routing mechanisms, the network distributes traffic solely based on hashes, making path engineering or fault-based scheduling difficult. Once a link fails, the convergence time and policies of the underlying protocols often fail to meet the need for rapid recovery.

[0003] More importantly, Ethernet inherently lacks centralized control capabilities. The network control plane and data plane are loosely coupled, and switches have limited awareness of each other, making it impossible to build a "global network view." This is particularly prominent in multi-tenant, hyper-converged, and RDMA over Ethernet (such as RoCE v2) scenarios. For example, RDMA has extremely high requirements for lossless operation, low latency, and path determinism, while current Ethernet systems struggle to detect congestion in real time, perform path switching, or adjust QoS policies. This limits its further implementation in AI and HPC scenarios.

[0004] OpenSM (Open Subnet Manager) is one of the core software components of the InfiniBand network, responsible for the management, path discovery, topology maintenance, and route configuration of the entire subnet. It communicates with the Subnet Management Agent (SMA) on each node by sending Subnet Management Packets (SMP), thereby managing all HCAs, switches, and routers within the subnet. During the topology discovery phase, OpenSM starts from the local management port and uses Directed Route SMP to probe the network hop-by-hop, sending management requests to each port in sequence to obtain the peer's GUID, port number, link status, and speed information, and constructing a complete node-link relationship graph. Once all ports and nodes have been probed, it internally generates a topology database for subsequent LID allocation and route calculation. Afterward, OpenSM periodically performs a Subnet Sweep to detect network changes, port online / offline status, or new devices, and performs incremental updates on the changed parts.

[0005] Topology discovery in SDN (Software-Defined Networking) is centrally handled by the controller. Its core objective is to establish a global view of the network, including all switch nodes, port status, link connections, and host access points. After establishing connections with each switch via the OpenFlow control channel, the controller periodically sends out specially crafted LLDP probe packets. When a neighboring switch receives these LLDP packets but has no matching flow table entry, it reports this to the controller as a Packet-In event. The controller parses the switch ID (DPID) and port information in the packets to determine the physical connection between the two ports, thereby constructing the network link topology. Simultaneously, the controller also identifies the host access location and port status based on the PortStatus events reported by the switches and the source ports of ordinary data packets (such as ARP and DHCP). All node, link, and host information is stored in the controller's internal topology database and dynamically updated through periodic LLDP probes and event triggering. Through this mechanism, the SDN controller can monitor the entire network structure in real time, providing a global decision-making basis for path calculation, flow table distribution, traffic engineering, and fault recovery.

[0006] OpenSM's topology discovery mechanism relies on hop-by-hop Directed Route SMP probing. In large-scale InfiniBand networks, this can lead to lengthy initial topology scans, especially in data centers with hundreds of switches and thousands of ports, where a single sweep process can last tens of seconds. It uses a centralized management model, with only one master SM responsible for network-wide probing and maintenance. If this process exits abnormally or becomes overloaded, topology updates for the entire subnet will stall. OpenSM handles dynamic changes (port online / offline, link drift) primarily through periodic sweep and trap events, resulting in weak real-time performance and a susceptibility to route recalculation or broadcast storms during high-frequency topology changes. Since SMP protocols are point-to-point queries, management traffic increases significantly as the network scales, placing additional burdens on bandwidth and management node CPUs.

[0007] SDN controllers rely on LLDP active probing and Packet-In passive reporting for topology discovery. While this approach is cross-platform and highly versatile, it has significant limitations. First, the periodic delivery of LLDP messages can overload the control plane. When the topology is large or the LLDP cycle is too short, the controller is prone to message congestion and link oscillations. Second, topology change detection relies on asynchronous event reporting (PortStatus, PacketIn), which can lead to inconsistencies or outdated information in high-latency or packet-loss environments. Furthermore, compatibility with heterogeneous devices from multiple vendors is limited. Most importantly, SDN's centralized control model makes topology discovery highly sensitive to controller availability and link stability; if the control channel is interrupted, network state awareness immediately fails. Summary of the Invention

[0008] In view of the above problems, the present invention is proposed to provide an adaptive network topology detection method for AI computing power data centers that overcomes or at least partially solves the above problems.

[0009] According to one aspect of the present invention, an adaptive network topology detection method based on AI computing power data centers is provided, the detection method comprising: Step S1: Start up the physical equipment in the data center and prepare for the connection of cables to ensure the physical link connection of the AI ​​computing power data center network. Step S2: Deploy an agent program that supports the LTDP protocol based on the LLDP protocol extension on the switching equipment of the computing power network data center; Step S3: Start the agent program on the switch device to collect local LTDP information, including LLDP information and network characteristic information related to interfaces, flow control, and devices. Step S4: Deploy the adaptive network topology detection device on the general server node of the computing power network data center and start it.

[0010] Optionally, the LTDP protocol, compared to the LLDP protocol, adds interface, flow control, and device-related computing power network characteristic information, specifically including: counting the number of ports under the agent, counting the port speed under the agent, counting the maximum transmission unit (MTU) of the port under the agent, counting the FIB information (Table size) under the switch, identifying the device type under the agent (Device-type), and the unique identifier of the device (Device-UI).

[0011] Optionally, step S4: deploying the adaptive network topology detection device on the general server node of the computing power network data center, and starting it specifically includes: Step S4.1: The adaptive network topology detection device constructs a TDP protocol message and sends it to the adjacent switch of the server; Step S4.2: The adjacent switch will respond to the received TDP message, verify the M_Key management key in the message, and send the message back to the adaptive network topology detection device; Step S4.3: The adaptive network topology detection device receives the TDP message returned by the adjacent switch, parses and stores the Data information in the message, continues to construct TDP messages to detect the next-hop adjacent switch device of the adjacent switch, and sets the Hop Count in the message to 2. Step S4.4: After receiving the probe packet, the adjacent switch finds that the Hop Count value in the packet is 2 and the Hop Pointer value is 1, and continues to send the packet to the next-hop adjacent switch. Step S4.5: After receiving the packet, the next-hop neighboring switch of the adjacent switch finds that the Hop Count value is 2 and the HopPointer value is 2. It determines that the packet requires a response from this device, verifies the M_Key management key in the packet, constructs a response packet, and sends the packet back to the adaptive network topology detection device. Step S4.6: The adaptive network topology detection device receives the returned TDP message, parses and stores the Data information in the message, continues to construct TDP messages to detect the remote switch device, and sets the Hop Count in the message to 3; This process continues until the entire topology collection algorithm converges.

[0012] Optionally, constructing the response message in step S4.1 includes: setting D to 1 for return, setting R to 1 for response, setting the Method operation method to GetResp, and filling in the Data to collect local LTDP information based on the LLDP protocol extension.

[0013] Optionally, the adaptive network topology detection device constructs a TDP protocol message and sends it to the adjacent switch of the server. The field names and meanings in the TDP protocol message include: D indicates message direction: 0 indicates outbound and 1 indicates inbound; Status method returns status code; Hop Pointer points to the current hop position in Initial Path; Hop Count indicates the number of valid bytes in Initial / Return Path, which is the number of path hops; M_Key manages the key used to verify the main program's permissions. The configuration operation must match: DrSUID (source LID, i.e., the main program server port identifier), DrDUID (target LID, used to indicate the last-hop device), Data (method parameter data area, the specific content of which is determined by the Method), Initial Path (each byte of the initial path is the device name and port number), and Return Path (return path). Each byte is filled back by the intermediate device as the device name and port number for the reverse jump, the R response bit: 0 = request, 1 = response, and the Method operation method.

[0014] Optionally, the construction of the response message in step S4.5 includes: D set to 1 for return, R response position set to 1, Method operation method is GetResp, and Data is filled with collected local LTDP information based on the LLDP protocol extension.

[0015] Optionally, the specific content of step S4.1 includes: reading node information and writing port configuration.

[0016] This invention provides an adaptive network topology detection method for AI computing power data centers. The detection method includes: Step S1: Starting up the physical equipment of the data center and preparing the cabling to ensure the physical links of the AI ​​computing power data center network are connected; Step S2: Deploying an agent program based on the LTDP protocol extension (LLDP protocol extension) on the switching equipment of the computing power network data center; Step S3: Starting the agent program on the switching equipment to collect local LTDP information based on the LLDP protocol extension, including LLDP information and computing power network characteristic information related to interfaces, flow control, and devices; Step S4: Deploying the adaptive network topology detection device on the general server nodes of the computing power network data center and starting it. Using the TDP protocol algorithm, the adaptive network topology detection device sends TDP data to adjacent switching devices, and the adjacent devices respond with LTDP data and perform hop-by-hop forwarding, ultimately realizing intelligent detection and information collection of network device topology within the computing power center.

[0017] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating an adaptive network topology detection method based on an AI computing power data center, provided as an embodiment of the present invention; Figure 2 This is a schematic diagram of the detection process of the adaptive network topology detection device provided in an embodiment of the present invention; Figure 3 This is a network topology diagram of an AI computing power data center provided in an embodiment of the present invention; Figure 4 Explanation of each field of the TDP message provided in this embodiment of the invention; Figure 5 Explanation of the fields of Method provided in the embodiments of the present invention. Detailed Implementation

[0020] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0021] The terms "comprising" and "having," and any variations thereof, in the specification, embodiments, claims, and drawings of this invention are intended to cover non-exclusive inclusion, such as including a series of steps or units.

[0022] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0023] like Figure 1 As shown, this invention provides an adaptive network topology detection method based on AI computing power data centers. The detection method includes: Step S1: Start-up of data center physical equipment and preparation for cable connectivity, ensuring the physical links of AI computing power data center network equipment are connected, such as... Figure 3 As shown; Step S2: Deploy an agent program based on the LTDP protocol, an extension of the LLDP protocol, on the switching equipment of the computing power network data center; Step S3: Start the agent program on the switch device to collect local LTDP information based on the LLDP protocol extension, including LLDP information and interface, flow control, and device-related computing power network characteristic information; LLDPDU is the data body, composed of multiple TLVs, and supports extensible field encapsulation. TLV is the unit that makes up LLDPDU, and each TLV represents a piece of information. For LLDPDU, the newly added TLV contains the following information: Port number (counts the number of ports under this agent), Port speed (counts the port speed under this agent), Port mtu (counts the maximum transmission unit of the port under this agent), Table size (counts the FIB information under this switch), Device-type (identifies the device type under this agent), Device-UI (composed of sysname and mac, used as a unique identifier for the device).

[0024] Step S4: Deploy the adaptive network topology detection device on the general server node of the computing power network data center and start it.

[0025] Step S4.1: The adaptive network topology detection device constructs a TDP (Topology Discovery Packet) protocol message and sends it to the adjacent switch of the server. The message contains the following settings: Hop Count is set to 1, D (Direction Bit) is set to 0 to indicate outbound traffic, DrDUID (Target LID) is set to the broadcast bit, and R (Response Bit) is set to 0. The message format and explanation are as follows: Figure 4 and Figure 5 As shown. The detection process of the adaptive network topology detection device is as follows. Figure 2 As shown.

[0026] Figure 4 Explanation of each field in the TDP message; Figure 5 Explanation of each field in Method.

[0027] Step S4.2: The adjacent switch responds to the received TDP message, verifies the M_Key management key in the message, constructs a response message including: D (Direction Bit) set to 1 for return, R response position set to 1, Method operation method is GetResp, Data is filled with collected local LTDP information based on the LLDP protocol extension, and sends the message back to the adaptive network topology detection device.

[0028] Step S4.3: The adaptive network topology detection device receives the TDP message returned by the adjacent switch, parses and stores the Data information in the message, continues to construct TDP messages to detect the next-hop adjacent switch device of the adjacent switch, and sets the Hop Count in the message to 2.

[0029] Step S4.4: After receiving the probe message, the adjacent switch finds that the Hop Count value in the message is 2 and the HopPointer value is 1, and continues to send the message to the next-hop adjacent switch.

[0030] Step S4.5: After receiving the packet, the next-hop adjacent switch of the adjacent switch finds that the Hop Count value is 2 and the HopPointer value is 2. It determines that the packet requires a response from this device, verifies the M_Key management key in the packet, constructs a response packet including: D (Direction Bit) set to 1 for return, R response position set to 1, Method operation method is GetResp, Data is filled with collected local LTDP information based on the LLDP protocol extension, and sends the packet back to the adaptive network topology detection device.

[0031] Step S4.6: The adaptive network topology detection device receives the returned TDP message, parses and stores the Data information in the message, and continues to construct TDP messages to probe remote switch devices, setting the Hop Count in the message to 3. This process continues until the entire topology collection algorithm converges.

[0032] Beneficial Effects: Traditional AI computing power data center network topology information collection relies on configuring LLDP on each switching device, with the controller connecting to each switch via GRPC to collect LLDP information. This requires manual configuration of each device and lacks computing power information collection. This adaptive network topology detection method for AI computing power data centers uses a combination of LTDP and TDP protocols. After each switching device collects topology neighbor information, the topology detection device automatically detects and collects topology information from all switching device nodes by constructing TDP probe messages and sending them hop-by-hop, reducing manual intervention.

[0033] After traditional data center networking is completed, if there are abnormal transmission links or equipment failures, maintenance personnel need to troubleshoot. This method automatically detects faults by the device or neighboring devices and actively reports relevant fault information to the detection device using the TDP protocol, triggering the detection device to perform a re-topology scan and topology information management.

[0034] The InfiniBand network computing power data center controller OpenSM subnet probing method collects probing packets by broadcasting them node by node, gradually probing each port of each node. However, this method suffers from slow convergence due to the large number of forwarded packets. This approach uses a combination of LTDP and TDP protocols. Each switch uses LTDP to pre-collect local neighbor information, while the probing device uses TDP for probing, resulting in faster convergence and reduced packet forwarding.

[0035] The above specific embodiments further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An adaptive network topology detection method based on AI computing power data centers, characterized in that, The detection method includes: Step S1: Start up the physical equipment in the data center and prepare for the connection of cables to ensure the physical link connection of the AI ​​computing power data center network. Step S2: Deploy an agent program that supports the LTDP protocol based on the LLDP protocol extension on the switching equipment of the computing power network data center; Step S3: Start the agent program on the switch device to collect local LTDP information, including LLDP information and network characteristic information related to interfaces, flow control, and devices. Step S4: Deploy the adaptive network topology detection device on the general server node of the computing power network data center and start it.

2. The adaptive network topology detection method based on AI computing power data center according to claim 1, characterized in that, Compared to the LLDP protocol, the LTDP protocol adds interface, flow control, and device-related computing power network characteristic information, specifically including: the number of ports under the agent, the port speed under the agent, the maximum transmission unit (mtu) of the port under the agent, the FIB information table size under the switch, the device type (Device-type) under the agent, and the unique identifier (Device-UI) of the device.

3. The adaptive network topology detection method based on AI computing power data center according to claim 1, characterized in that, Step S4: Deploying the adaptive network topology detection device on the general server node of the computing power network data center, and starting it specifically includes: Step S4.1: The adaptive network topology detection device constructs a TDP protocol message and sends it to the adjacent switch of the server; Step S4.2: The adjacent switch will respond to the received TDP message, verify the M_Key management key in the message, and send the message back to the adaptive network topology detection device; Step S4.3: The adaptive network topology detection device receives the TDP message returned by the adjacent switch, parses and stores the Data information in the message, continues to construct TDP messages to detect the next-hop adjacent switch device of the adjacent switch, and sets the Hop Count in the message to 2. Step S4.4: After receiving the probe packet, the adjacent switch finds that the Hop Count value in the packet is 2 and the Hop Pointer value is 1, and continues to send the packet to the next-hop adjacent switch. Step S4.5: After receiving the packet, the next-hop neighboring switch of the adjacent switch finds that the Hop Count value is 2 and the HopPointer value is 2. It determines that the packet requires a response from this device, verifies the M_Key management key in the packet, constructs a response packet, and sends the packet back to the adaptive network topology detection device. Step S4.6: The adaptive network topology detection device receives the returned TDP message, parses and stores the Data information in the message, continues to construct TDP messages to detect the remote switch device, and sets the Hop Count in the message to 3; This process continues until the entire topology collection algorithm converges.

4. The adaptive network topology detection method based on AI computing power data center according to claim 3, characterized in that, The construction of the response message in step S4.1 includes: setting D to 1 for return, setting R to 1 for response, setting the Method operation method to GetResp, and filling in the Data to collect local LTDP information based on the LLDP protocol extension.

5. The adaptive network topology detection method based on AI computing power data center according to claim 3, characterized in that, The adaptive network topology detection device constructs a TDP protocol message and sends it to the adjacent switch of the server. The field names and meanings in the TDP protocol message include: D indicates message direction: 0 indicates outbound and 1 indicates inbound; Status method returns status code; Hop Pointer points to the current hop position in Initial Path; Hop Count indicates the number of valid bytes in Initial / Return Path, which is the number of path hops; M_Key manages the key used to verify the main program's permissions. The configuration operation must match: DrSUID (source LID, i.e., the main program server port identifier), DrDUID (target LID, used to indicate the last-hop device), Data (method parameter data area, the specific content of which is determined by the Method), Initial Path (each byte of the initial path is the device name and port number), and Return Path (return path). Each byte is filled back by the intermediate device as the device name and port number for the reverse jump, the R response bit: 0 = request, 1 = response, and the Method operation method.

6. The adaptive network topology detection method based on AI computing power data center according to claim 3, characterized in that, The construction of the response message in step S4.5 includes: D set to 1 for return, R response position set to 1, Method operation method is GetResp, and Data is filled with collected LTDP information based on the LLDP protocol extension.

7. The adaptive network topology detection method based on AI computing power data center according to claim 3, characterized in that, The specific content of step S4.1 includes: reading node information and writing port configuration.

Images