AI intelligent computing center management method, device and system based on automatic topology discovery
By using automatic topology discovery and remote API calls, the network topology of the AI computing center can be automatically updated and managed in a unified manner. This solves the problem of network topology relying on manual configuration in existing technologies, improves resource utilization and task execution efficiency, reduces operation and maintenance complexity, and supports dynamic changes of heterogeneous devices.
Patent Information
- Application Number
- CN202511184132.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-22
- Publication Date
- 2025-10-31
AI Technical Summary
The existing AI computing center relies on manual configuration for network topology discovery, which cannot automatically detect changes in network structure. This results in inflexible resource scheduling, a disconnect between task scheduling and network status, complex management of heterogeneous devices, a lack of efficient distributed communication mechanisms, and poor scalability.
By periodically collecting network neighbor information, the system automatically constructs and maintains the physical network topology, generates the optimal on-network computing topology tree, and uses remote call interfaces to collect device status, issue commands, and schedule tasks. Combined with streaming communication, a unified protocol framework, and intelligent topology management, it achieves automatic device discovery, unified management, and dynamic task scheduling.
It greatly improves network resource utilization, significantly reduces operation and maintenance complexity, enhances the execution efficiency and reliability of large model training tasks, has strong scalability and adaptability, and supports dynamic changes in heterogeneous devices and network structures.
Smart Images
Figure CN120880919A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to an AI intelligent computing center management method, apparatus, and system based on automatic topology discovery. Background Technology
[0002] An AI computing center is a complete system based on AI computer clusters built with AI chips, encompassing infrastructure (data center infrastructure), hardware infrastructure, and software infrastructure. It is primarily used for AI deep learning model development, training, and inference, providing full-stack AI capabilities from underlying chip computing power release to top-level application enablement. Currently, the network controllers of AI computing centers are mainly managed through static configuration and manual methods, including:
[0003] 1. Manual static configuration of network topology
[0004] Network administrators manually configure parameters such as IP addresses, VLANs, routing tables, and ACLs for devices such as switches, servers, and routers.
[0005] 2. Script-based batch configuration
[0006] Partial automation can be achieved by batch issuing configuration commands using scripts such as Shell and Python.
[0007] 3. Traditional SDN controllers (such as OpenDaylight, ONOS)
[0008] Network flow tables are centrally managed using protocols such as OpenFlow, enabling a certain degree of automated traffic scheduling.
[0009] 4. Vendor-owned network management system
[0010] Examples include Cisco DNA Center and Huawei iMaster NCE, which provide graphical interfaces and automated configuration functions.
[0011] 5. Configuration management based on static templates
[0012] Network device parameters can be managed uniformly and distributed in batches by configuring templates (such as YAML and JSON).
[0013] 6. Manual monitoring and troubleshooting
[0014] It relies on manual monitoring of network status and manual handling of issues such as link failures and equipment malfunctions.
[0015] The drawback of the above solution is:
[0016] Network topology discovery relies on manual configuration and cannot automatically detect changes in network structure, resulting in inflexible resource scheduling and a high risk of errors.
[0017] Task scheduling is disconnected from network status, making it impossible to dynamically optimize the distribution and execution paths of on-network computing tasks such as large model training based on the actual network topology and bandwidth conditions.
[0018] Multi-device heterogeneous management is complex, as the management interfaces of devices such as GPU servers and switches are not standardized, making it difficult to achieve automated and intelligent unified scheduling.
[0019] The lack of an efficient distributed communication mechanism results in low efficiency and poor scalability in operations such as task distribution and state synchronization. Summary of the Invention
[0020] This invention provides an AI intelligent computing center management method, apparatus, and system based on automatic topology discovery to improve the above-mentioned problems.
[0021] This invention provides a management method for AI intelligent computing centers based on automatic topology discovery, comprising:
[0022] It periodically collects network neighbor information through network topology protocols, and automatically constructs and maintains the physical network topology of the AI computing center based on the collected network neighbor information;
[0023] When receiving a task related to a large model, the system automatically generates the optimal on-network computing topology tree by combining the current physical network topology.
[0024] The task is decomposed according to the network computing topology tree, and the decomposed subtasks are sent to the device nodes; wherein, remote procedure calls are performed through remote call interfaces to realize the status collection of device nodes, command issuance, and task scheduling.
[0025] During the execution of subtasks on device nodes, the tasks are managed throughout their entire lifecycle, including task startup, monitoring, anomaly recovery, and result collection.
[0026] Preferably, the device node includes a device node GPU server and a switch; the remote call interface adopts the gRPC interface, and the device node runs a gRPC Agent to expose a unified management interface.
[0027] Preferably, network neighbor information is collected periodically through a network topology protocol, and the physical network topology of the AI computing center is automatically constructed and maintained based on the collected network neighbor information, specifically including:
[0028] The system periodically sends discovery requests to all device nodes through the network topology protocol, so that each device node can broadcast neighbor information through the network topology protocol and report the collected neighbor information after receiving the discovery request.
[0029] After aggregating all neighbor information, the physical network topology is built or updated.
[0030] Preferably, the network topology protocol is LLDP, SNMP, or Netconf.
[0031] Preferably, upon receiving a task related to a large model, automatically generating the optimal on-network computing topology tree based on the current physical network topology specifically includes:
[0032] Receive large model training tasks submitted by users;
[0033] Analyze task requirements; these requirements include the bandwidth, latency, and number of nodes required by the task.
[0034] The optimal on-network computing topology tree is generated based on task requirements and physical network topology; the algorithm for generating the on-network computing topology tree adopts different optimization objectives according to actual needs.
[0035] Preferably, the task is decomposed according to the network computing topology tree, and the decomposed subtasks are distributed to the device nodes, specifically including:
[0036] The task is broken down into subtasks, and the subtasks are assigned to appropriate device nodes;
[0037] Subtasks and flow table configurations are distributed to device nodes via remote calls, enabling each device node to execute on-network computing tasks based on the received subtasks.
[0038] During execution, monitor for any execution anomalies; if so, trigger fault recovery or rescheduling; otherwise, collect task results.
[0039] Preferably, during the task decomposition process, the task allocation strategy is determined by evaluating computational complexity, communication patterns, and predicting resource requirements.
[0040] During the subtask allocation process, a hierarchical resource allocator is used for allocation, including: allocation within GPU servers, allocation between GPU servers, and allocation via the switch engine.
[0041] During subtask scheduling, a multi-strategy scheduler is used for task scheduling, including FIFO scheduling, priority scheduling, and ML optimal scheduling.
[0042] This invention also provides an AI intelligent computing center management device based on automatic topology discovery, comprising:
[0043] The physical network topology building unit is used to periodically collect network neighbor information through network topology protocols, and automatically build and maintain the physical network topology of the AI intelligent computing center based on the collected network neighbor information;
[0044] The topology tree generation unit is used to automatically generate the optimal on-network computing topology tree when receiving tasks related to large models, in combination with the current physical network topology.
[0045] The task decomposition unit is used to decompose the task according to the network computing topology tree and distribute the decomposed subtasks to the device nodes; wherein, remote procedure calls are performed through remote call interfaces to realize the status collection, command issuance and task scheduling of device nodes;
[0046] The full lifecycle management unit is used to manage the entire lifecycle of tasks during the execution of subtasks on device nodes, including task startup, monitoring, anomaly recovery, and result collection.
[0047] This invention also provides an AI intelligent computing center management system based on automatic topology discovery, comprising:
[0048] The system comprises a central controller and multiple device nodes. The central controller performs remote process calls via a remote procedure call interface to achieve status acquisition, command issuance, and task scheduling of the device nodes.
[0049] The central controller is used to periodically collect network neighbor information through network topology protocols, and automatically build and maintain the physical network topology of the AI computing center based on the collected network neighbor information;
[0050] When receiving a task related to a large model, the system automatically generates the optimal on-network computing topology tree by combining the current physical network topology.
[0051] The task is decomposed according to the network computing topology tree, and the decomposed subtasks are sent to the device nodes;
[0052] A device node is configured to execute the subtask upon receiving it.
[0053] The central controller is also used to manage the entire lifecycle of tasks during the execution of subtasks on device nodes, including task initiation, monitoring, anomaly recovery, and result collection.
[0054] In summary, the embodiments of the present invention, through innovative technologies such as streaming communication, unified protocol framework, intelligent topology management, ancestor node path optimization, and breadth-first resource allocation, achieve automatic discovery and real-time updating of network topology in heterogeneous network environments, unified distributed device management and scheduling, dynamic on-network computing topology tree generation based on large model training requirements, and efficient task scheduling and lifecycle management.
[0055] Compared with the prior art, this embodiment has the following advantages:
[0056] 1. Significantly improves network resource utilization, achieving deep coupling and dynamic optimization between tasks and network status;
[0057] 2. Significantly reduces operation and maintenance complexity, enabling automatic device discovery, automatic registration, and unified management;
[0058] 3. Improve the execution efficiency and reliability of on-network computing tasks such as large model training;
[0059] 4. It has strong scalability and adaptability, supporting dynamic changes in heterogeneous devices and network structures. Attached Figure Description
[0060] Figure 1 This is a flowchart illustrating the AI intelligent computing center management method based on automatic topology discovery according to the first embodiment of the present invention.
[0061] Figure 2 A flowchart illustrating the process of automatic network topology discovery;
[0062] Figure 3 A flowchart illustrating the process of task decomposition, allocation, and scheduling;
[0063] Figure 4 A schematic diagram illustrating the principle of task decomposition and scheduling;
[0064] Figure 5 This is a schematic diagram of the structure of the AI intelligent computing center management device based on automatic topology discovery according to the second embodiment of the present invention. Detailed Implementation
[0065] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.
[0066] Please see Figure 1 The first embodiment of the present invention provides an AI intelligent computing center management method based on automatic topology discovery, which can be executed by an AI intelligent computing center management device (such as a central controller) based on automatic topology discovery to achieve the following steps:
[0067] S101 periodically collects network neighbor information through network topology protocols, and automatically constructs and maintains the physical network topology of the AI intelligent computing center based on the collected network neighbor information.
[0068] In this embodiment, the central controller and each device node perform remote process calls through a remote procedure call interface to achieve functions such as status acquisition, command issuance, and task scheduling of device nodes. It also supports automatic registration, heartbeat detection, and status synchronization of device nodes.
[0069] The remote call protocol can be gRPC, RESTful API, WebSocket, etc., but this invention does not specify a particular protocol. This embodiment preferably uses the gRPC protocol.
[0070] The device nodes may include heterogeneous devices such as GPU servers and switches. Each device node runs a gRPC Agent to expose a unified management interface.
[0071] In this embodiment, the network topology protocol can be LLDP, SNMP, or Netconf, etc., and the present invention does not impose a specific limitation. This embodiment preferably uses the LLDP protocol.
[0072] In this embodiment, specifically, as Figure 2 As shown, the central controller ( Figure 2 The INC Manager controller periodically sends discovery requests to all device nodes via the LLDP protocol. After receiving the discovery request, each device node broadcasts neighbor information via the network topology protocol. Each device node can collect the neighbor information broadcast by other device nodes and report the collected neighbor information to the central controller. The central controller then aggregates all neighbor information to construct the physical network topology (INC Network).
[0073] The central controller, after constructing the physical network topology, compares it with the current physical network topology and determines whether there has been a change. If there is no change, the existing topology is maintained; if there is a change, the network topology database is updated, and other modules are notified that the network topology has been updated, thereby realizing real-time perception and dynamic updating of network topology changes.
[0074] S102, upon receiving a task related to a large model, automatically generates the optimal on-network computing topology tree by combining the current physical network topology.
[0075] In this embodiment, specifically, after receiving a large model training task submitted by the user, the central controller analyzes the task requirements; the task requirements include the bandwidth, latency, and number of nodes required by the task.
[0076] Then, based on the task requirements and the physical network topology, an optimal in-network computing topology tree (INC Tree) is generated; the process of generating the optimal in-network computing topology tree is as follows:
[0077] First, calculations are performed based on the lowest common ancestor. Then, shortest path enumeration is performed to obtain candidate topology trees. Finally, the candidate topology trees are optimized to obtain the optimal on-network computation topology tree.
[0078] The optimization process employs different optimization objectives based on actual needs, as shown in Table 1.
[0079] Table 1: INC Tree Optimization Strategy
[0080] Optimization Objective Algorithm Strategy Applicable Scenarios Minimum Spanning Tree Minimize total hops / latency Delay-sensitive tasks Bandwidth Optimal Tree Maximize bottleneck bandwidth Big data transmission Load balancing tree Minimize load variance High-concurrency computing Fault-tolerant optimization tree Maximize path redundancy High availability requirements
[0081] S103, the task is decomposed according to the network computing topology tree, and the decomposed subtasks are sent to the device nodes; wherein, remote procedure calls are performed through remote call interfaces to realize the status collection, command issuance, and task scheduling of device nodes.
[0082] S104 performs full lifecycle management of tasks during the execution of subtasks on device nodes, including task startup, monitoring, anomaly recovery, and result collection.
[0083] Specifically, such as Figure 3 As shown:
[0084] First, the task is broken down into subtasks, and the subtasks are assigned to appropriate device nodes;
[0085] Then, subtasks and flow table configurations are sent to device nodes via remote calls, so that each device node can execute on-network computing tasks according to the received subtasks;
[0086] Finally, during execution, monitor for any execution anomalies; if so, trigger fault recovery or rescheduling; if not, collect task results.
[0087] Preferably, such as Figure 4 As shown, during the task decomposition process, the task allocation strategy is determined by evaluating computational complexity, communication patterns, and predicting resource requirements.
[0088] During the subtask allocation process, a hierarchical resource allocator is used for allocation, including: allocation within GPU servers, allocation between GPU servers, and allocation via the switch engine.
[0089] During subtask scheduling, a multi-strategy scheduler is used for task scheduling, including FIFO scheduling, priority scheduling, and ML optimal scheduling.
[0090] In summary, the embodiments of the present invention, through innovative technologies such as streaming communication, unified protocol framework, intelligent topology management, ancestor node path optimization, and breadth-first resource allocation, achieve automatic discovery and real-time updating of network topology in heterogeneous network environments, unified distributed device management and scheduling, dynamic on-network computing topology tree generation based on large model training requirements, and efficient task scheduling and lifecycle management.
[0091] Compared with the prior art, this embodiment has the following advantages:
[0092] 1. Significantly improves network resource utilization, achieving deep coupling and dynamic optimization between tasks and network status;
[0093] 2. Significantly reduces operation and maintenance complexity, enabling automatic device discovery, automatic registration, and unified management;
[0094] 3. Improve the execution efficiency and reliability of on-network computing tasks such as large model training;
[0095] 4. It has strong scalability and adaptability, supporting dynamic changes in heterogeneous devices and network structures.
[0096] Please see Figure 5 The second embodiment of the present invention also provides an AI intelligent computing center management device based on automatic topology discovery, comprising:
[0097] The physical network topology construction unit 210 is used to periodically collect network neighbor information through network topology protocols, and automatically construct and maintain the physical network topology of the AI intelligent computing center based on the collected network neighbor information;
[0098] The topology tree generation unit 220 is used to automatically generate the optimal on-network computing topology tree when receiving tasks related to large models, in combination with the current physical network topology.
[0099] The task decomposition unit 230 is used to decompose the task according to the network computing topology tree and distribute the decomposed subtasks to the device nodes; wherein, remote procedure calls are performed through a remote procedure call interface to realize the status acquisition, command issuance and task scheduling of the device nodes;
[0100] The full lifecycle management unit 240 is used to manage the task throughout its entire lifecycle during the execution of subtasks on the device node, including task startup, monitoring, anomaly recovery, and result collection.
[0101] The third embodiment of the present invention also provides an AI intelligent computing center management system based on automatic topology discovery, comprising:
[0102] The system comprises a central controller and multiple device nodes. The central controller performs remote process calls via a remote procedure call interface to achieve status acquisition, command issuance, and task scheduling of the device nodes.
[0103] The central controller is used to periodically collect network neighbor information through network topology protocols, and automatically build and maintain the physical network topology of the AI computing center based on the collected network neighbor information;
[0104] When receiving a task related to a large model, the system automatically generates the optimal on-network computing topology tree by combining the current physical network topology.
[0105] The task is decomposed according to the network computing topology tree, and the decomposed subtasks are sent to the device nodes;
[0106] A device node is configured to execute the subtask upon receiving it.
[0107] The central controller is also used to manage the entire lifecycle of tasks during the execution of subtasks on device nodes, including task initiation, monitoring, anomaly recovery, and result collection.
[0108] In the several embodiments provided in this invention, it should be understood that the disclosed methods can also be implemented in other ways. The apparatus and method embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, 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 a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive 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 a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.
[0109] In addition, the functional modules in the various embodiments of the present invention can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.
[0110] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, electronic device, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks. It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. In the absence of further restrictions, an element defined by the phrase "comprising a..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0111] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. 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. A management method for AI intelligent computing centers based on automatic topology discovery, characterized in that, include: It periodically collects network neighbor information through network topology protocols, and automatically constructs and maintains the physical network topology of the AI computing center based on the collected network neighbor information; When receiving a task related to a large model, the system automatically generates the optimal on-network computing topology tree by combining the current physical network topology. The task is decomposed according to the network computing topology tree, and the decomposed subtasks are sent to the device nodes; wherein, remote procedure calls are performed through remote call interfaces to realize the status collection of device nodes, command issuance, and task scheduling. During the execution of subtasks on device nodes, the tasks are managed throughout their entire lifecycle, including task startup, monitoring, anomaly recovery, and result collection.
2. The AI intelligent computing center management method based on automatic topology discovery according to claim 1, characterized in that, The device node includes a device node GPU server and a switch; the remote call interface adopts the gRPC interface, and the device node runs a gRPC Agent to expose a unified management interface.
3. The AI intelligent computing center management method based on automatic topology discovery according to claim 1, characterized in that, The system periodically collects network neighbor information via network topology protocols, and automatically constructs and maintains the physical network topology of the AI computing center based on this information. Specifically, this includes: The system periodically sends discovery requests to all device nodes through the network topology protocol, so that each device node can broadcast neighbor information through the network topology protocol and report the collected neighbor information after receiving the discovery request. After aggregating all neighbor information, the physical network topology is built or updated.
4. The AI intelligent computing center management method based on automatic topology discovery according to claim 1, characterized in that, The network topology protocol is LLDP, SNMP, or Netconf.
5. The AI intelligent computing center management method based on automatic topology discovery according to claim 1, characterized in that, When receiving tasks related to large models, the system automatically generates the optimal on-network computing topology tree based on the current physical network topology, specifically including: Receive large model training tasks submitted by users; Analyze task requirements; these requirements include the bandwidth, latency, and number of nodes required by the task. The optimal on-network computing topology tree is generated based on task requirements and physical network topology; the algorithm for generating the on-network computing topology tree adopts different optimization objectives according to actual needs.
6. The AI intelligent computing center management method based on automatic topology discovery according to claim 2, characterized in that, The task is decomposed according to the network computing topology tree, and the decomposed subtasks are distributed to the device nodes, specifically including: The task is broken down into subtasks, and the subtasks are assigned to appropriate device nodes; Subtasks and flow table configurations are distributed to device nodes via remote calls, enabling each device node to execute on-network computing tasks based on the received subtasks. During execution, monitor for any execution anomalies; if so, trigger fault recovery or rescheduling; otherwise, collect task results.
7. The AI intelligent computing center management method based on automatic topology discovery according to claim 6, characterized in that, During the task decomposition process, the task allocation strategy is determined by evaluating computational complexity, communication patterns, and predicting resource requirements. During the subtask allocation process, a hierarchical resource allocator is used for allocation, including: allocation within GPU servers, allocation between GPU servers, and allocation via the switch engine. During subtask scheduling, a multi-strategy scheduler is used for task scheduling, including FIFO scheduling, priority scheduling, and ML optimal scheduling.
8. An AI-powered intelligent computing center management device based on automatic topology discovery, characterized in that, include: The physical network topology building unit is used to periodically collect network neighbor information through network topology protocols, and automatically build and maintain the physical network topology of the AI intelligent computing center based on the collected network neighbor information; The topology tree generation unit is used to automatically generate the optimal on-network computing topology tree when receiving tasks related to large models, in combination with the current physical network topology. The task decomposition unit is used to decompose the task according to the network computing topology tree and distribute the decomposed subtasks to the device nodes; wherein, remote procedure calls are performed through remote call interfaces to realize the status collection, command issuance and task scheduling of device nodes; The full lifecycle management unit is used to manage the entire lifecycle of tasks during the execution of subtasks on device nodes, including task startup, monitoring, anomaly recovery, and result collection.
9. An AI-powered intelligent computing center management system based on automatic topology discovery, characterized in that, include: The system comprises a central controller and multiple device nodes. The central controller performs remote process calls via a remote procedure call interface to achieve status acquisition, command issuance, and task scheduling of the device nodes. The central controller is used to periodically collect network neighbor information through network topology protocols, and automatically build and maintain the physical network topology of the AI computing center based on the collected network neighbor information; When receiving a task related to a large model, the system automatically generates the optimal on-network computing topology tree by combining the current physical network topology. The task is decomposed according to the network computing topology tree, and the decomposed subtasks are sent to the device nodes; A device node is configured to execute the subtask upon receiving it. The central controller is also used to manage the entire lifecycle of tasks during the execution of subtasks on device nodes, including task initiation, monitoring, anomaly recovery, and result collection.