Simulation method and platform of multi-machine multi-card AI computing cluster and storage medium

Abstract

The invention discloses a simulation method and platform of a multi-machine multi-card AI computing cluster and a storage medium. The simulation method comprises the following steps: generating global network configuration information based on a network topology structure input to a simulation platform; in response to a communication operation request of an AI calculation acceleration unit, sending an inter-unit request based on the global network configuration information; and in response to the received inter-unit request, based on address information in the inter-unit request and local network configuration information acquired from the global network configuration information, distributing the inter-unit request to a specified AI calculation acceleration unit for processing.

Description

Technical Field

[0001] This invention generally relates to the fields of computer simulation and distributed computing technology, and more specifically, to a simulation method, simulation platform, and computer storage medium for a multi-machine, multi-card artificial intelligence (AI) computing cluster. Background Technology

[0002] As large-scale models with hundreds of billions of parameters become the mainstream of AI development, their training and inference tasks rely on computing clusters consisting of hundreds or thousands of servers equipped with multiple AI computing acceleration units. Developing, debugging, and optimizing system software (such as deep learning frameworks and ensemble communication libraries) on such multi-machine, multi-GPU clusters faces enormous challenges: real-world cluster hardware costs are extremely high, making them unsuitable for frequent destructive testing or in-depth performance profiling; and the defects and performance bottlenecks of distributed software are highly dependent on specific network environments, making them difficult to reproduce and locate stably in real-world clusters. Summary of the Invention

[0003] This application provides a simulation method, platform, computer storage medium, and computer program product for a multi-machine, multi-GPU AI computing cluster. It aims to build a simulation environment that is compatible with the real AI software ecosystem, supports fine-grained hardware modeling, and provides in-depth performance insights through software-defined virtualization technology, thereby significantly reducing the development, testing, and optimization costs and timelines of AI distributed system software.

[0004] In this article, "inter-unit request" refers to a message or instruction simulated in the simulation environment, which is sent by one AI computing acceleration unit (through the inter-unit client of its node) to another AI computing acceleration unit (through the inter-unit server of the target node) for communication and interaction.

[0005] According to one aspect of this application, a simulation method for a multi-machine, multi-GPU AI computing cluster is provided, comprising: generating global network configuration information based on a network topology input to a simulation platform; sending an inter-unit request based on the global network configuration information in response to a communication operation request from an AI computing acceleration unit; and distributing the inter-unit request to a designated AI computing acceleration unit for processing based on address information in the inter-unit request and local network configuration information obtained from the global network configuration information in response to receiving the inter-unit request.

[0006] According to another aspect of this application, a simulation platform for a multi-machine, multi-GPU AI computing cluster is provided, comprising: a network setting module configured to generate global network configuration information based on a network topology input to the simulation platform; an inter-unit client configured to send an inter-unit request based on the global network configuration information in response to a communication operation request from an AI computing acceleration unit; and an inter-unit server configured to distribute the inter-unit request to a designated AI computing acceleration unit for processing based on address information in the inter-unit request and local network configuration information obtained from the global network configuration information in response to receiving the inter-unit request.

[0007] According to another aspect of this application, a computer-readable storage medium is provided, storing instructions that, when executed by one or more processors, cause the one or more processors to perform the method described above.

[0008] According to another aspect of this application, a computer program product is provided, including instructions that, when executed by one or more processors, cause the one or more processors to perform the method described above. Attached Figure Description

[0009] Figure 1 A flowchart of a multi-machine, multi-GPU AI computing cluster simulation method according to this application is illustrated as an example.

[0010] Figure 2 An exemplary schematic diagram of the structure of a simulation platform for a multi-machine, multi-GPU AI computing cluster according to this application is shown.

[0011] Figure 3 An exemplary embodiment of the method for implementing this application is shown. Figure 1 A schematic diagram of the simulation device for the method shown. Detailed Implementation

[0012] The features and exemplary embodiments of various aspects of the present invention will now be described in detail. Numerous specific details are set forth in the following detailed description in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention may be practiced without requiring some of these specific details. The following description of embodiments is merely intended to provide a better understanding of the invention by illustrating examples of the invention. The invention is by no means limited to any specific configurations and algorithms presented below, but covers any modifications, substitutions, and improvements to elements, components, and algorithms without departing from the spirit of the invention. Well-known structures and techniques are not shown in the drawings and the following description in order to avoid unnecessarily obscuring the invention.

[0013] It should be noted that the "AI computing acceleration unit" mentioned in this application refers to a hardware component with AI-specific computing and communication capabilities, including but not limited to graphics processing units (GPUs), tensor processing units (TPUs), field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), and neural network processors (NPUs). Among them, "AI computing acceleration unit" refers to the entity that initiates the communication operation request (i.e., the communication initiator AI computing acceleration unit), and "designated AI computing acceleration unit" refers to the target entity that receives and processes the request between units (i.e., the communication receiver AI computing acceleration unit). The two are clearly distinguished by the division of roles in the communication process, and will not be described separately in the following description.

[0014] The technical solution of this application achieves simulation through global network configuration and inter-unit communication scheduling. Its adaptability covers various AI computing acceleration units. Adaptation to different types of acceleration units does not require changes to the core technical process; only the call interception method of the underlying communication library and the configuration parameters of the hardware interconnection topology within the node need to be adjusted according to their hardware characteristics to achieve the corresponding communication simulation. The invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention.

[0015] Figure 1 A flowchart of a multi-machine, multi-GPU AI computing cluster simulation method according to this application is illustrated by way of example. In one or more embodiments, the method is... Figure 2 The simulation is executed on the simulation platform 200 shown. The simulation platform 200 includes at least: a network setup module 202, inter-unit clients 204a and 204b, and inter-unit servers 206a and 206b. In one or more embodiments, each virtual node may have an inter-unit client and an inter-unit server deployed on it. For example, virtual node A has an inter-unit client 204a and an inter-unit server 206a deployed on it, and is configured with at least one AI computing acceleration unit 208a; virtual node B has an inter-unit client 204b and an inter-unit server 206b deployed on it, and is configured with at least one AI computing acceleration unit 208b.

[0016] In one or more embodiments, a single inter-unit communication process involves two roles located on different virtual nodes: an inter-unit client 204a on the initiating virtual node (e.g., virtual node A) sends a request to an inter-unit server 206b on the target virtual node (e.g., virtual node B). Therefore, for any given inter-unit communication, the client and server are always deployed on different virtual nodes. The inter-unit client is configured to initiate communication requests on behalf of its local virtual node to the target inter-unit server located on other virtual nodes; the inter-unit server is configured to receive and process communication requests from other virtual nodes on behalf of its own virtual node.

[0017] like Figure 1 As shown, the method 100 may include: Step S102: Based on the network topology input to the simulation platform, generate global network configuration information. In one or more embodiments, this step is performed by... Figure 2 The network settings module 202 is executed. In one or more embodiments, the network topology describes the physical or logical connection relationships of the target multi-machine, multi-card AI cluster, including the number of nodes, network connections between nodes (such as InfiniBand switch topology), etc.

[0018] In one or more embodiments, the global network configuration information includes intra-node hardware interconnect topology information (e.g., hardware interconnect topology between AI computing acceleration units (such as a connection matrix), and the connection relationship between AI computing acceleration units and network interface cards (NICs)) and link simulation parameters. The link simulation parameters include at least one of bandwidth, latency, and packet loss rate, used to simulate real network behavior in subsequent steps.

[0019] Global network configuration and routing information provide inter-unit clients with the necessary information to address the target inter-unit server, thereby defining the feasible paths and methods for inter-unit communication between any two virtual nodes in the simulation environment.

[0020] In one or more embodiments, the generated global network configuration information and routing information are stored in a centralized configuration service or distributed to various components within the platform for their querying and use. The global network configuration information also includes, or can be used to generate, routing policies between nodes for client addressing between units.

[0021] Step S104: In response to the communication operation request from the AI ​​computing acceleration unit, an inter-unit request is sent based on global network configuration information. In one or more embodiments, this step is performed by... Figure 2 The client executes 204a or 204b between the middle units.

[0022] Reference Figure 2 In response to a communication operation request from AI computing acceleration unit 208a, inter-unit client 204a in virtual node A queries global network configuration information, determines the target virtual node B to which the communication should be delivered based on the routing policy therein, and then sends an inter-unit request to the target inter-unit server 206b deployed on that target node. Similarly, in response to a communication operation request from AI computing acceleration unit 208b, inter-unit client 204b in virtual node B queries global network configuration information, determines the target virtual node A to which the communication should be delivered based on the routing policy therein, and then sends an inter-unit request to the target inter-unit server 206a deployed on that target node.

[0023] In one or more embodiments, the inter-unit client responding to a communication operation request includes: obtaining the AI ​​computing acceleration unit's call instruction to the underlying hardware communication library, and using the call instruction as a communication operation request. Specifically, API calls to the underlying hardware communication library by the process where the AI ​​computing acceleration unit resides can be intercepted and collected through library function hijacking (e.g., using the LD_PRELOAD mechanism), thereby transparently capturing the communication intent of the unmodified AI application.

[0024] In one or more embodiments, during the process of sending inter-unit requests, the inter-unit clients can simulate network transmission behavior based on link simulation parameters in the global network configuration information. For example, they can inject waiting time into the sending thread according to the latency parameter to simulate transmission delay; randomly decide whether to drop requests according to the packet loss rate parameter to simulate network packet loss; or control the packet sending rate according to the bandwidth parameter.

[0025] In one or more embodiments, the inter-unit client can also dynamically select a communication algorithm model (such as selecting a P2P model for small messages and a Tree algorithm for large messages) based on the message size of the inter-unit request and the hardware interconnection topology information within the node, and simulate the communication process according to the steps of the selected model, thereby more accurately reflecting the behavior of the real communication library under a specific hardware topology.

[0026] Step S106: In response to receiving an inter-unit request, based on the address information in the inter-unit request and the local network configuration information obtained from the global network configuration information, the inter-unit request is distributed to the designated AI computing acceleration unit for processing.

[0027] Reference Figure 2In virtual node A, the target unit inter-server 206a, upon receiving an inter-unit request from the inter-unit client 204b in virtual node B, performs a routing decision based on the address information (such as the target AI computing acceleration unit ID) in the inter-unit request and the local network configuration information matching this virtual node obtained from the global network configuration information, to determine which specific AI computing acceleration unit 208a within this node should process the request. Similarly, in virtual node B, the target unit inter-server 206b, upon receiving an inter-unit request from the inter-unit client 204a in virtual node A, performs a routing decision based on the address information (such as the target AI computing acceleration unit ID) in the inter-unit request and the local network configuration information matching this virtual node obtained from the global network configuration information, to determine which specific AI computing acceleration unit 208b within this node should process the request.

[0028] The above cross-node communication scenarios are merely illustrative examples. The technical solutions of this application are not limited to the interaction of AI computing acceleration units between different virtual nodes, but can also be flexibly adapted to the communication needs within the same node.

[0029] Specifically, the cross-node communication scenario illustrated in Figure 2 is merely for the purpose of understanding the technical solution of this application and is not intended to limit the scope of protection. The simulation method and platform of this application are also applicable to simulation scenarios of communication between multiple AI computing acceleration units within the same node: when the communication initiator and receiver are different AI computing acceleration units within the same virtual node, the hardware interconnection topology information within the node in the global network configuration information will provide a routing basis for the transmission of requests between units. In this simulation scenario of communication within the same node, the logical instances of the client and the server between units can be deployed within the same virtual node process and complete the request transmission through the local communication channel without the need for cross-node network link forwarding. Moreover, the link simulation parameters can be configured to adapt to the hardware interconnection characteristics within the node (such as low latency and high bandwidth), thereby accurately simulating the communication behavior and resource competition status of multiple AI computing acceleration units within the same node.

[0030] In one or more embodiments, the inter-unit server can prioritize multiple inter-unit requests based on the load status of the multiple AI computing acceleration units it manages, and then distribute the requests to the corresponding AI computing acceleration units in sequence according to the scheduling results, so as to simulate the competition for computing resources within a node.

[0031] The distributed inter-unit requests are "processed" by the designated AI computing acceleration unit according to its built-in computing model, simulating real data computing or caching operations, thereby completing a complete cross-node communication simulation closed loop.

[0032] In one or more embodiments, the simulation method of this application may further include an analysis and optimization phase. The simulation platform 200 can collect various performance data (such as latency, throughput, and computing unit utilization) during the sending, distribution, and processing of requests between units in steps S104 and S106, and generate a performance analysis report based on this data. In one or more embodiments, the simulation platform 200 can automatically adjust the global network configuration information or the operating parameters of the AI ​​computing acceleration unit based on this report, and drive the simulation process to re-execute, achieving closed-loop automatic tuning, thereby providing data-driven decision support for AI cluster configuration and job parameter optimization.

[0033] It should be noted that, Figure 2 The diagram illustrates only one exemplary deployment and communication scenario. In actual simulations, each virtual node, as shown in Virtual Node A and Virtual Node B, can be equipped with its own inter-unit client and inter-unit server. The path shown in the diagram from the client of Virtual Node A to the server of Virtual Node B is only used to illustrate the interaction between the client and server in cross-node communication.

[0034] This application enables the construction of an AI cluster "digital twin" on a general-purpose server, fully compatible with the production environment software stack and supporting everything from chip interconnects to fine-grained global network modeling. Developers can directly run real distributed training jobs within it, debug communication logic without loss, reproduce network-related defects, and proactively evaluate system performance under different hardware architectures or software parameters.

[0035] Compared with existing technologies, this invention significantly reduces the development and debugging costs and risks of AI distributed systems, improves software quality and performance optimization efficiency, and has important value for promoting the development of AI infrastructure technology.

[0036] The simulation method and platform provided in this application can be widely used in AI chip design companies, cloud service providers, large Internet companies and research institutions for AI hardware architecture evaluation, development and testing of distributed system software (communication library, scheduler), cluster configuration selection, and performance prediction and parameter tuning of production operations. It has significant industrial practicality and economic value.

[0037] Figure 3 An exemplary embodiment of the method for implementing this application is shown. Figure 1 A schematic diagram of the simulation apparatus for the method shown is presented, which is a computing device 300 that can be used to execute the method described above. Figure 3 As shown, computing device 300 may include bus 302 or other communication mechanism for transmitting information, and one or more hardware processors 304 coupled to bus 302 for processing information. The one or more hardware processors 304 may include, for example, one or more general-purpose microprocessors.

[0038] like Figure 3 As shown, in some embodiments, computing device 300 may further include main memory 306 coupled to bus 302. Main memory 306 is used to store information and instructions executed by one or more processors 304, such as random access memory (RAM), cache, and / or other dynamic storage devices. Main memory 306 may also be used to store temporary variables or other intermediate information during the execution of instructions executed by one or more processors 304. When these instructions are stored in storage media accessible to one or more processors 304, they can cause computing device 300 to become a dedicated machine customized to perform the operations specified in the instructions. Storage device 308 may include non-volatile and / or volatile storage media. Non-volatile storage media may include, for example, optical discs or magnetic disks. Volatile storage media may include dynamic memory. Common forms of storage media may include, for example, floppy disks, hard disks, solid-state drives, magnetic tape, or any other magnetic data storage media, CD-ROMs, any other optical data storage media, any physical media with a perforated pattern, RAM, DRAM, PROM, and EPROM, FLASH-EPROM, NVRAM, any other memory chip or cartridge, or their networking versions.

[0039] like Figure 3 As shown, in some embodiments, computing device 300 may further include one or more communication interfaces or network interfaces 310 coupled to bus 302. Network interface 310 may provide bidirectional data communication coupling to one or more network links connected to one or more networks. As another example, network interface 310 may be a local area network (LAN) card to provide data communication connectivity to a LAN-compatible (or WAN component communicating with a WAN) network. Wireless links may also be implemented.

[0040] The execution of certain operations can be distributed across processors rather than residing within a single machine, but rather deployed across multiple machines. In some example embodiments, the processor or processor-implemented engine may reside in a single geographic location (e.g., in a home environment, office environment, or server farm). In other example embodiments, the processor or processor-implemented engine may be distributed across multiple geographic locations.

[0041] Each of the processes, methods, and algorithms described in the preceding sections may be embodied in code modules executed by one or more computer systems or computer processors including computer hardware, and may be fully or partially automated by these code modules. The processes and algorithms may be implemented, partially or fully, in dedicated circuit systems.

[0042] When the functions disclosed herein are implemented as software functional units and sold or used as stand-alone products, they may be stored in a processor-executable, non-volatile, computer-readable storage medium. Specific technical solutions (all or part) disclosed herein, or aspects contributing to the prior art, may be embodied in the form of a software product. The software product may be stored in a storage medium and includes several instructions that cause a computing device (which may be a personal computer, server, network device, etc.) to perform all or some steps of the methods of the embodiments of this application. The storage medium may include a flash drive, portable hard disk drive, ROM, RAM, magnetic disk, optical disk, other media operable to store program code, or any combination thereof.

[0043] Specific embodiments further provide a system including a processor and a non-transitory computer-readable storage medium storing instructions executable by the processor to cause the system to perform operations corresponding to steps in any method of the embodiments disclosed above. Specific embodiments further provide a non-transitory computer-readable storage medium storing instructions executable by one or more processors to cause the one or more processors to perform operations corresponding to steps in any method of the embodiments disclosed above.

[0044] The embodiments disclosed herein can be implemented via a cloud platform, server, or server cluster (collectively referred to below as the "Service System") that interacts with a client. The client can be a terminal device or a client registered by a user at the platform, wherein the terminal device can be a mobile terminal, a personal computer (PC), or any device that can have the platform application installed.

[0045] The various features and processes described above can be used independently of each other or combined in various ways. All possible combinations and sub-combinations are intended to fall within the scope of this disclosure. Additionally, certain method or process blocks may be omitted in some embodiments. The methods and processes described herein are not limited to any particular order, and their associated blocks or states may be executed in other suitable orders. For example, described blocks or states may be executed in an order other than that specifically disclosed, or multiple blocks or states may be combined into a single block or state. Example blocks or states may be executed sequentially, in parallel, or in some other manner. Blocks or states may be added to or removed from the disclosed example embodiments. The exemplary systems and components described herein may be configured differently than described. For example, components may be added to, removed from, or rearranged compared to the disclosed example embodiments.

[0046] The various operations of the exemplary methods described herein can be performed at least in part by an algorithm. The algorithm may be included in program code or instructions stored in memory (e.g., the aforementioned non-transitory computer-readable storage medium). This algorithm may include a machine learning algorithm. In some embodiments, the machine learning algorithm may not explicitly refer to the computer as performing the function but may learn from training data to generate a predictive model of the function.

[0047] The various operations of the exemplary methods described herein can be performed, at least in part, by one or more processors that are temporarily configured (e.g., by software) or permanently configured to perform the relevant operations. Whether temporarily or permanently configured, these processors can constitute an engine of processor implementation that operates to perform one or more of the operations or functions described herein.

[0048] Similarly, the methods described herein may be implemented at least in part by a processor, wherein one or more specific processors are instances of hardware. For example, at least some operations of the methods may be performed by one or more processors or an engine implemented by a processor. Furthermore, one or more processors may also be operable to support the execution of relevant operations in a “cloud computing” environment or as the execution of relevant operations in a “Software as a Service” (SaaS) context. For example, at least some operations may be performed by a group of computers (as an example of a machine containing processors), wherein these operations are accessible via a network (e.g., the Internet) and via one or more appropriate interfaces (e.g., application programming interfaces (APIs)).

[0049] The execution of certain operations can be distributed across processors rather than residing within a single machine, and can be deployed across multiple machines. In some example embodiments, the processor or processor-implemented engine may reside in a single geographic location (e.g., in a home environment, office environment, or server farm). In other example embodiments, the processor or processor-implemented engine may be distributed across multiple geographic locations.

[0050] Throughout this specification, multiple instances may be implemented as components, operations, or structures described as a single instance. Although individual operations of one or more methods are illustrated and described as separate operations, one or more of these individual operations may be performed simultaneously, and not necessarily in the order illustrated. Structures and functions presented as separate components in the example configuration may be implemented as composite structures or components. Similarly, structures and functions presented as single components may be implemented as single components. These and other variations, modifications, additions, and improvements fall within the scope of this document.

[0051] As used herein, "or" is inclusive rather than exclusive unless explicitly indicated by the context. Therefore, in this document, "A, B, or C" means "A, B, A and B, A and C, B and C, or A, B, and C" unless explicitly indicated by the context. Furthermore, "and" is combined and separate unless explicitly indicated by the context. Therefore, in this document, "A and B" means "A and B, combined or separate" unless explicitly indicated by the context. Additionally, multiple instances of resources, operations, or structures described herein may be provided as a single instance. Furthermore, the boundaries between various resources, operations, engines, and data storage devices are somewhat arbitrary and specific operations are illustrated within the context of a particular illustrative configuration. Other functional assignments are foreseeable and fall within the scope of various embodiments of this disclosure. Generally, structures and functions presented as individual resources in example configurations may be implemented as combined structures or resources. Similarly, structures and functions presented as single resources may be implemented as single resources. These and other changes, modifications, additions, and improvements fall within the scope of the embodiments of this disclosure as expressed in the appended claims. Therefore, this specification and drawings should be considered illustrative rather than restrictive.

[0052] The terms “comprising” or “including” are used to indicate the presence of a subsequently claimed feature, but do not preclude the addition of other features. Unless otherwise specifically stated or otherwise understood in the context in which they are used, conditional language such as “may,” “can,” “may,” and “can” is generally intended to convey that certain embodiments include certain features, components, and / or steps that are not included in other embodiments. Therefore, this conditional language is generally not intended to imply that one or more embodiments require features, components, and / or steps in any way, or that one or more embodiments must include logic for determining whether such features, components, and / or steps are included in or performed in any particular embodiment, with or without user input or prompts.

[0053] Although the general outline of the subject matter has been described with reference to specific exemplary embodiments, various modifications and changes may be made to these embodiments without departing from the broad scope of embodiments of this disclosure. Where more than one embodiment is disclosed, these embodiments of the subject matter may be referred to individually or collectively herein as the term "invention," this is for convenience only and is not intended to automatically limit the scope of this application to any single disclosure or concept.

[0054] The embodiments illustrated herein are described in detail to enable those skilled in the art to practice the disclosed teachings. Other embodiments may be used and derived therefrom, such that structural and logical substitutions and changes may be made without departing from the scope of this disclosure. Therefore, “implementation” is not intended to be limiting, and the scope of the various embodiments is defined only by the appended claims and their full scope.

Claims

1. A simulation method for a multi-machine, multi-GPU AI computing cluster, characterized in that, The simulation method includes: Based on the network topology input to the simulation platform, global network configuration information is generated. In response to a communication operation request from the AI ​​computing acceleration unit, an inter-unit request is sent based on the global network configuration information; and In response to receiving the inter-unit request, based on the address information in the inter-unit request and the local network configuration information obtained from the global network configuration information, the inter-unit request is distributed to the designated AI computing acceleration unit for processing.

2. The method according to claim 1, characterized in that, The global network configuration information includes intra-node hardware interconnection topology information and link simulation parameters.

3. The method according to claim 2, characterized in that, The link simulation parameters include at least one of bandwidth, latency, and packet loss rate.

4. The method according to claim 1, characterized in that, Responding to the communication operation request includes: obtaining the AI ​​computing acceleration unit's call instruction to the underlying hardware communication library, and using the call instruction as the communication operation request.

5. The method according to claim 4, characterized in that, Obtaining the call instruction includes: intercepting the API calls of the underlying hardware communication library by the process where the AI ​​computing acceleration unit is located through library function hijacking.

6. The method according to claim 1, characterized in that, Distributing the inter-unit requests to designated AI computing acceleration units for processing includes: scheduling multiple received inter-unit requests based on the load status of the AI ​​computing acceleration units, and distributing the inter-unit requests to the corresponding AI computing acceleration units in sequence according to the scheduling results.

7. The method according to claim 2, characterized in that, The inter-unit request includes: simulating network transmission behavior based on the link simulation parameters.

8. The method according to claim 7, characterized in that, The simulated network transmission behavior includes at least one of the following: simulating transmission waiting based on the delay parameter in the link simulation parameters, simulating packet loss events based on the packet loss rate parameter, and simulating transmission rate limitation based on the bandwidth parameter.

9. The method according to claim 2, characterized in that, The process of sending inter-unit requests further includes: selecting a communication algorithm model based on the message size of the inter-unit request and the hardware interconnection topology information within the node, and simulating the corresponding inter-unit communication according to the selected communication algorithm model.

10. The method according to any one of claims 1 to 9, characterized in that, Also includes: Collect performance data during the processing of requests between sending units, requests between distribution units, and the AI ​​computing acceleration unit; A performance analysis report is generated based on the performance data.

11. The method according to claim 10, characterized in that, Also includes: Based on the performance analysis report, adjust the global network configuration information or the operating parameters of the AI ​​computing acceleration unit, and re-execute the simulation process of the multi-machine, multi-card AI computing cluster.

12. A simulation platform for a multi-machine, multi-GPU AI computing cluster, characterized in that, The simulation platform includes: The network settings module is configured to generate global network configuration information based on the network topology input to the simulation platform. Inter-unit clients are configured to respond to communication operation requests from AI computing acceleration units by sending inter-unit requests based on the global network configuration information; and The inter-unit server is configured to, in response to receiving the inter-unit request, distribute the inter-unit request to a designated AI computing acceleration unit for processing based on the address information in the inter-unit request and the local network configuration information obtained from the global network configuration information.

13. The simulation platform according to claim 12, characterized in that, The inter-unit server and the inter-unit client that initiates the request are deployed on different virtual nodes. The local network configuration information is the network configuration information that matches the virtual node where the inter-unit server is located. The designated AI computing acceleration unit is the AI ​​computing acceleration unit deployed on the virtual node where the inter-unit server is located.

14. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed by one or more processors, cause the one or more processors to perform a simulation method for a multi-machine, multi-GPU AI computing cluster according to any one of claims 1 to 11.

Images