Artificial intelligence computing method, device, electronic device and medium

Abstract

This specification provides an artificial intelligence computing method, device, electronic device and medium. The method includes: generating a network demand report according to a collective communication log corresponding to the artificial intelligence computing, the network demand report records the demand of the collective communication for the network; detecting the network status of the network according to the network demand report to obtain the detection result; optimizing the topology of the network according to the detection result; and performing artificial intelligence computing based on the optimized network.

Description

Technical Field

[0001] This specification relates to the field of artificial intelligence technology, and in particular to artificial intelligence computing methods, devices, electronic devices and media. Background Art

[0002] In artificial intelligence training, collective communication is a global communication operation in which all processes in a process group participate. Its most basic operations include send, receive, copy, intra-group process barrier synchronization Barrier, and inter-node process synchronization (signal + wait). These most basic operations are combined to form a set of communication templates, also called communication primitives, such as: 1-to-many broadcast, many-to-1 gather, many-to-many all-gather, 1-to-many scatter, many-to-1 reduce, many-to-many all-reduce, combined reduce and divergent reduce-scatter, many-to-many all-to-all, etc.

[0003] The network is an external component that is called when the collective communication topology algorithm needs to send traffic, and serves as a module for underlying traffic transmission. This design easily leads to a disconnect between the network configuration and the collective communication topology algorithm. Although the network is normal, the collective communication often fails to work. Summary of the invention

[0004] To overcome the problems existing in the related art, this specification provides an artificial intelligence computing method, device, electronic device and medium.

[0005] According to a first aspect of an embodiment of the present specification, there is provided an artificial intelligence computing method, the method comprising: generating a network demand report according to a collective communication log corresponding to the artificial intelligence computing, the network demand report recording the demand of the collective communication for the network; detecting the network status of the network according to the network demand report to obtain a detection result; optimizing the topology of the network according to the detection result; and performing artificial intelligence computing based on the optimized network.

[0006] According to the second aspect of the embodiments of this specification, there is provided an artificial intelligence computing device, including: a report generation module, used to generate a network demand report according to a collective communication log corresponding to the artificial intelligence computing, wherein the network demand report records the demand of the collective communication for the network; a network detection module, used to detect the network status of the network according to the network demand report and obtain a detection result; an optimization module, used to optimize the topology of the network according to the detection result; and a computing module, used to perform artificial intelligence computing based on the optimized network.

[0007] According to a third aspect of the embodiments of this specification, there is provided an electronic device, including:

[0008] processor;

[0009] a memory for storing processor-executable instructions;

[0010] Among them, the processor is configured to execute the artificial intelligence computing method of the above-mentioned first aspect or any corresponding embodiment thereof.

[0011] According to a fourth aspect of the embodiments of this specification, a computer-readable storage medium is provided, on which computer instructions are stored, and the computer instructions are used to enable a computer to execute the artificial intelligence computing method of the above-mentioned first aspect or any corresponding embodiment thereof.

[0012] The technical solutions provided by the embodiments of this specification may have the following beneficial effects:

[0013] In the embodiments of this specification, the requirements of the collective communication topology algorithm for the network are extracted, and it is automatically detected whether the current network meets the requirements. If not, the network topology structure is optimized so that the collective communication network meets the requirements of artificial intelligence computing, thereby improving the stability of the collective communication network and the reliability of artificial intelligence computing.

[0014] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present specification. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the specification and, together with the description, serve to explain the principles of the specification.

[0016] Figure 1 It is a schematic diagram of a system architecture shown in this specification according to an exemplary embodiment.

[0017] Figure 2 It is a flowchart of an artificial intelligence computing method shown in this specification according to an exemplary embodiment.

[0018] Figure 3A It is a schematic diagram of an NCCL log according to an exemplary embodiment of this specification.

[0019] Figure 3B It is a schematic diagram of an NCCL log according to another exemplary embodiment of the present specification.

[0020] Figure 3CIt is a schematic diagram of an NCCL log according to another exemplary embodiment of the present specification.

[0021] Figure 3D It is a schematic diagram of basic information of a network card according to an exemplary embodiment of this specification.

[0022] Figure 4 It is a hardware structure diagram of the computer device where the artificial intelligence computing device of the embodiment of this specification is located.

[0023] Figure 5 It is a block diagram of an artificial intelligence computing device shown in this specification according to an exemplary embodiment. DETAILED DESCRIPTION

[0024] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementations described in the following exemplary embodiments do not represent all implementations consistent with this specification. Instead, they are merely examples of devices and methods consistent with some aspects of this specification as detailed in the appended claims.

[0025] The terms used in this specification are for the purpose of describing specific embodiments only and are not intended to limit this specification. The singular forms "a", "the" and "the" used in this specification and the appended claims are also intended to include plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more associated listed items.

[0026] It should be understood that although the terms first, second, third, etc. may be used in this specification to describe various information, this information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of this specification, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0027] Next, the embodiments of this specification are described in detail.

[0028] The following combination Figure 1 The system architecture to which the artificial intelligence computing method and device can be applied in the embodiments of this specification is described. It should be noted that: Figure 1What is shown is merely an example of a system architecture to which the embodiments of this specification can be applied, to help those skilled in the art understand the technical content of this specification, but it does not mean that the embodiments of this specification cannot be used in other devices, systems, environments or scenarios.

[0029] Figure 1 It is a schematic diagram of a system architecture shown in this specification according to an exemplary embodiment.

[0030] like Figure 1 As shown, the system architecture may include, for example, a terminal device, a network and a server. The network is used to provide a medium for a communication link between the terminal device and the server. The network may include various connection types, such as wired and / or wireless communication links, etc.

[0031] Users can use terminal devices to interact with servers through the network to receive or send messages, etc. Various communication client applications can be installed on the terminal devices, such as artificial intelligence applications, web browser applications, search applications, instant messaging tools, email clients and / or social platform software.

[0032] The terminal device may be any electronic device having a display screen and supporting web browsing, including but not limited to a smart phone, a tablet computer, a laptop computer, a desktop computer, and the like.

[0033] The server can be a server that provides various services, such as a background management server that provides support for the content browsed by the user using the terminal device. The background management server can analyze and process the received user request and other data, and feed back the processing results (such as web pages, information, or data obtained or generated according to the user request) to the terminal device.

[0034] The artificial intelligence calculation method provided in the embodiments of this specification is described in detail below. Figure 2 As shown, Figure 2 This is a flowchart of an artificial intelligence computing method shown in this specification according to an exemplary embodiment. The method can be applied to a server, for example. The artificial intelligence computing method provided in the embodiment of this specification may include the following steps.

[0035] In step 210, a network demand report is generated based on the aggregate communication log corresponding to the artificial intelligence calculation.

[0036] According to an embodiment of the present specification, a network demand report records the demand for a network by collective communication, and may include, for example, an access subject, access direction, access method, access network card, etc., that needs to access the network in collective communication. Access subjects may include source subjects and destination subjects. Access directions may include, for example, sending and receiving. An access method may, for example, represent a method used when accessing a network. An access network card may, for example, represent a network card used when accessing a network. Collective communication logs may, for example, include NCCL (NVIDIA Collective Communications Library) logs.

[0037] According to an embodiment of the present specification, for example, the correspondence information between the graphics processor GPU and the host node can be determined based on the usage device information in the collective communication log. According to the communication path information in the collective communication log, the source GPU, the destination GPU, the access method, the access direction, and the access network card are determined. According to the correspondence information, the source host node corresponding to the source GPU and the destination host node corresponding to the destination GPU are determined. In addition, the basic information of the access network card is queried. Then, a network demand report is generated based on the source GPU, the source host node, the destination GPU, the destination host node, the access method, the access direction, the access network card, and the basic information of the access network card.

[0038] In step 220, the network status of the network is detected according to the network demand report to obtain a detection result.

[0039] According to an embodiment of the present specification, for example, by detecting the network status of the network, an abnormal node in the network can be determined as a detection result.

[0040] In step 230, the topology of the network is optimized according to the detection result.

[0041] According to an embodiment of the present specification, for example, an abnormal node may be deleted from the topology structure of the network to recalculate the topology structure of the network.

[0042] In step 240, artificial intelligence calculations are performed based on the optimized network.

[0043] According to the embodiments of this specification, artificial intelligence computing may include operations such as training models, natural language processing, computer vision processing, knowledge representation and reasoning, etc.

[0044] According to the embodiments of this specification, the requirements of the collective communication topology algorithm for the network are extracted, and it is automatically detected whether the current network meets the requirements. If not, the network topology structure is optimized so that the collective communication network meets the requirements of artificial intelligence computing, thereby improving the stability of the collective communication network and the reliability of artificial intelligence computing.

[0045] Optionally, step 220 may include, for example: detecting whether each access network card in the network requirement report can communicate normally. If it is detected that at least one access network card in the network requirement report cannot communicate normally, the access network card that cannot communicate normally is recorded in the detection result.

[0046] According to the embodiments of the present specification, for example, for each access network card, the IP address of the outgoing interface can be searched in the routing table of the access network card; it is determined whether the IP address of the outgoing interface and the IP address of the corresponding gateway are in the same subnet; if the IP address of the outgoing interface and the IP address of the corresponding gateway are not in the same subnet, it is determined that the access network card cannot communicate normally. And / or, for each access network card, it can be detected whether the message sent by the access network card can reach the opposite network card; if the message sent by the access network card cannot reach the opposite network card, it is determined that the access network card cannot communicate normally.

[0047] Based on this, step 230 may include, for example: if the detection result includes an access network card that cannot communicate normally, deleting the access network card that cannot communicate normally from the topology structure.

[0048] Optionally, step 220 may include, for example: detecting whether the network parameters of each access network card in the network demand report are abnormal. If it is detected that the network parameters of at least one access network card in the network demand report are abnormal, the access network card with abnormal network parameters is recorded in the detection result.

[0049] According to an embodiment of the present specification, for example, for each access network card, the bandwidth and / or delay between the access network card and the peer network card can be detected. If the bandwidth and / or delay is within an abnormal range, it is determined that the network parameters of the access network card are abnormal. The abnormal range can be set according to actual needs.

[0050] Based on this, step 230 may include, for example: if the detection result includes an access network card with abnormal network parameters, deleting the access network card with abnormal network parameters from the topology structure.

[0051] Optionally, for example, the corresponding relationship information between the graphics processing GPU and the host node can be determined based on the device information used in the collective communication log, and the corresponding relationship information includes the number of the GPU participating in the collective communication in the collective communication, the host identifier of the host node where the GPU is located, and the device identifier of the GPU in the host node;

[0052] According to the communication path information in the collective communication log, the host ID of the source host, the number of the source GPU, the number of the destination GPU, the access method, the access direction, and the network card ID of the access network card are determined; according to the corresponding relationship information, the host ID of the destination host and the number of the destination GPU, the device ID of the destination GPU is determined; in addition, the basic information of the access network card can be obtained according to the network card ID; then, the network demand report can be determined according to the host ID of the source host, the number of the source GPU, the device ID of the source GPU, the host ID of the destination host, the number of the destination GPU, the device ID of the destination GPU, the access method, the access direction, the network card ID of the access network card and the basic information of the access network card.

[0053] The artificial intelligence calculation method is described below in conjunction with another exemplary embodiment. The artificial intelligence calculation method may include the following steps:

[0054] Obtain NCCL logs and analyze the network requirements of collective communication based on the logs.

[0055] Step 301: Read Using devices information from the log. The Using devices information may include the Rank number of the GPU participating in the collective communication, the group identifier Group, the process identifier Pid, ​​the host identifier of the host node where the GPU is located, and the device identifier device of the GPU in the host node.

[0056] The GPU number in the collective communication can be determined based on the Using devices information, and the corresponding relationship between the GPU and the host node can be analyzed and recorded.

[0057] Figure 3A It is a schematic diagram of an NCCL log according to an exemplary embodiment of this specification.

[0058] For example Figure 3A As shown in the figure, the NCCL log can include the Using devices information. The number after Rank is the number of the GPU in the topology algorithm, ai-k8s-node-ps-a800-gpu-5, ai-k8s-node-ps-a800-gpu-6, etc. are the IDs of the host node, and the number after device is the device ID of the GPU on the host node.

[0059] Step 302: First entries containing the first keyword can be filtered out from the NCCL log, and the requirements of the topology algorithm for the network can be extracted from these first entries to obtain a requirements report. The first keyword is used to filter the communication topology entries related to the network. The requirements report may include the access subject, access direction, access method, access network card, etc. of the network.

[0060] Figure 3B It is a schematic diagram of an NCCL log according to another exemplary embodiment of the present specification.

[0061] For example Figure 3B As shown in the figure, the first entry can be filtered out from the NCCL log according to the keyword "via NET", and the following access requirements can be extracted from the first entry:

[0062] Access subject: rank 5 (source GPU) of host node ai-k8s-node-ps-a800-gpu-5, rank 12 (destination GPU) of host node ai-k8s-node-ps-a800-gpu-6;

[0063] Access direction: receive (receive);

[0064] Access method: IB (InfiniBand, infinite bandwidth) network, and adopt GDR (GPUDirect RDMA, direct data transmission between GPU and remote direct memory access device) technology;

[0065] Access network card: Network card 6.

[0066] Step 303: A second entry containing a second keyword can be filtered out from the NCCL log, and the device identification of the network card can be extracted according to the second entry. The second keyword is used to filter all network cards in use. The second keyword corresponds to the access method in the demand report. According to the device identification, the basic information corresponding to the network card is queried, including the interface identification.

[0067] Figure 3C It is a schematic diagram of an NCCL log according to another exemplary embodiment of the present specification.

[0068] For example Figure 3C As shown, the device identification (mlx5_6:1) of the network card 6 can be confirmed by the second keyword "NCCL INFO NET / IB: Using", and the basic information corresponding to the network card 6 can be queried by the show_gids instruction and the device identification. The show_gids instruction can be used to output the basic information of the network card.

[0069] Figure 3D It is a schematic diagram of basic information of a network card according to another exemplary embodiment of the present specification.

[0070] For example Figure 3DAs shown, the basic information of the network card may include, for example, a global identifier (GID), an IP address, and an interface identifier, etc. For example, the basic information of the network card and other network information can be obtained at ai-k8s-node-ps-a800-gpu-5 and ai-k8s-node-ps-a800-gpu-6.

[0071] All the network requirements of the collective communication topology algorithm are sorted out to form a demand report and passed to the network status detection module. The demand report can be used to include the communication relationship between host nodes, as well as the network card, IP address and interface used by the host node during communication.

[0072] The network status detection module searches for the next-hop IP address of each network card in the policy routing table / routing table, and searches for the interface corresponding to the IP address in the gateway ARP address table for each network card in the demand report. It determines whether the interface is consistent with the interface corresponding to the network card in the demand report. If they are consistent, it proceeds to the next step. If they are inconsistent, it records the abnormal log.

[0073] For each network card in the demand report, check whether the link between the network card and the peer network card of the network card is normal.

[0074] For example, you can use ping to detect the reachability of the destination address, specify the source address as the address of the local network card, and the destination address as the address of the peer network card. If the access is normal, proceed to the next step. If the access is not possible, record the exception log.

[0075] For each pair of host nodes with communication relationship in the demand report, enable the server of the test tool perftest on the destination node in the pair of host nodes, and enable the client of perftest on the source node in the pair of host nodes. Based on the server and client of perftest, perform IB network connectivity detection. If the status is normal, record the bandwidth and / or latency results. If the status is abnormal, record the abnormal log.

[0076] Organize the network status detection results and send them to the report generation module.

[0077] The report generation module generates reports, organizes the network requirements of the communication topology and the network status detection results, outputs the reports, and provides them to users for analysis and processing.

[0078] For training optimization, use the environment variable NCCL_IB_HCA to remove network cards that cannot communicate and / or have abnormal parameters (such as low bandwidth rate) from the collective communication topology calculation, and restart the topology calculation for training.

[0079] Corresponding to the embodiments of the aforementioned methods, this specification also provides embodiments of an artificial intelligence computing device and a terminal to which it is applied.

[0080] The embodiments of the artificial intelligence computing device in this specification can be applied to computer devices, such as servers or terminal devices. The device embodiments can be implemented through software, hardware, or a combination of software and hardware. Taking software implementation as an example, as a device in a logical sense, it is formed by the processor in which it is located reading the corresponding computer program instructions in the non-volatile memory into the memory and running them. From the hardware level, if Figure 4 The figure is a hardware structure diagram of the computer device where the artificial intelligence computing device of the embodiment of this specification is located, except Figure 4 In addition to the processor 410, memory 430, network interface 420, and non-volatile memory 440 shown, the server or electronic device where the device 431 is located in the embodiment may also include other hardware according to the actual function of the computer device, which will not be described in detail.

[0081] like Figure 5 As shown, Figure 5 is a block diagram of an artificial intelligence computing device according to an exemplary embodiment of the present specification, the device comprising:

[0082] A report generation module 510, configured to generate a network demand report based on the collective communication log corresponding to the artificial intelligence calculation, wherein the network demand report records the demand of the collective communication for the network;

[0083] The network detection module 520 is used to detect the network status of the network according to the network demand report and obtain the detection result;

[0084] The optimization module 530 is used to optimize the topology of the network according to the detection result;

[0085] The computing module 540 is used to perform artificial intelligence computing based on the optimized network.

[0086] Optionally, the report generation module may include:

[0087] The relationship determination submodule is used to determine the corresponding relationship information between the graphics processor GPU and the host node according to the use device information in the collective communication log;

[0088] The demand determination submodule is used to determine the source GPU, the destination GPU, the access mode, the access direction, and the access network card according to the communication path information in the collective communication log;

[0089] A host determination submodule, used to determine a source host node corresponding to the source GPU and a destination host node corresponding to the destination GPU according to the corresponding relationship information;

[0090] The query submodule is used to query the basic information of the access network card;

[0091] The generation submodule is used to generate a network demand report according to the source GPU, the source host node, the destination GPU, the destination host node, the access mode, the access direction, the access network card and the basic information of the access network card.

[0092] Optionally, the network detection module may include:

[0093] The first detection submodule is used to detect whether each access network card in the network demand report can communicate normally;

[0094] A first recording submodule is used for recording the access network card that cannot communicate normally in the detection result if it is detected that at least one access network card in the network demand report cannot communicate normally;

[0095] Correspondingly, the optimization module may include:

[0096] The first deleting submodule is used to delete the access network card that cannot communicate normally from the topology structure if the detection result includes the access network card that cannot communicate normally.

[0097] Optionally, the network detection module may include:

[0098] The second detection submodule is used to detect whether the network parameters of each access network card in the network demand report are abnormal;

[0099] A second recording submodule is used for recording the access network card with abnormal network parameters in the detection result if it is detected that the network parameters of at least one access network card in the network demand report are abnormal;

[0100] Correspondingly, the optimization module may include:

[0101] The second deleting submodule is used to delete the access network card with abnormal network parameters from the topology structure if the detection result includes the access network card with abnormal network parameters.

[0102] Optionally, the first detection submodule may be specifically used for:

[0103] For each access network card, search the IP address of the outgoing interface in the routing table of the access network card; determine whether the IP address of the outgoing interface and the IP address of the corresponding gateway are in the same subnet; if the IP address of the outgoing interface and the IP address of the corresponding gateway are not in the same subnet, determine that the access network card cannot communicate normally; and / or

[0104] For each access network card, it is detected whether the message sent by the access network card can reach the opposite network card; if the message sent by the access network card cannot reach the opposite network card, it is determined that the access network card cannot communicate normally.

[0105] Optionally, the second detection submodule may be specifically used for:

[0106] For each access network card, the bandwidth and / or delay between the access network card and the opposite network card is detected; if the bandwidth and / or delay is within an abnormal range, it is determined that the network parameters of the access network card are abnormal.

[0107] According to the embodiments of this specification, the requirements of the collective communication topology algorithm for the network are extracted, and it is automatically detected whether the current network meets the requirements. If not, the network topology structure is optimized so that the collective communication network meets the requirements of artificial intelligence computing, thereby improving the stability of the collective communication network and the reliability of artificial intelligence computing.

[0108] Accordingly, the present specification also provides an electronic device, which includes a processor; a memory for storing instructions executable by the processor; wherein the processor is configured to: generate a network demand report based on a collective communication log corresponding to artificial intelligence calculations, the network demand report recording the demand of collective communication for the network; detect the network status of the network based on the network demand report to obtain detection results; optimize the topological structure of the network based on the detection results; and perform artificial intelligence calculations based on the optimized network.

[0109] The implementation process of the functions and effects of each module in the above-mentioned device is specifically described in the implementation process of the corresponding steps in the above-mentioned method, which will not be repeated here.

[0110] For the device embodiment, since it basically corresponds to the method embodiment, the relevant parts can refer to the partial description of the method embodiment. The device embodiment described above is only schematic, wherein the modules described as separate components may or may not be physically separated, and the components displayed as modules may or may not be physical modules, that is, they may be located in one place, or they may be distributed on multiple network modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this specification. A person of ordinary skill in the art can understand and implement it without paying creative labor.

[0111] The above is a description of a specific embodiment of the specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0112] Those skilled in the art will readily appreciate other embodiments of the specification after considering the specification and practicing the invention claimed herein. The specification is intended to cover any variations, uses or adaptations of the specification that follow the general principles of the specification and include common knowledge or customary techniques in the art that are not claimed in the specification. The specification and examples are to be considered exemplary only, and the true scope and spirit of the specification are indicated by the following claims.

[0113] It should be understood that the present description is not limited to the precise structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present description is limited only by the appended claims.

[0114] The above description is only a preferred embodiment of this specification and is not intended to limit this specification. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of this specification should be included in the scope of protection of this specification.

Claims

1. An artificial intelligence computing method, characterized in that: The method comprises: Generate a network demand report based on the collective communication log corresponding to the artificial intelligence calculation, wherein the network demand report records the demand of the collective communication for the network; According to the network demand report, detecting the network status of the network and obtaining a detection result; Optimizing the topology of the network according to the detection result; Perform artificial intelligence calculations based on the optimized network; The generating a network demand report for collective communication according to the collective communication log corresponding to the artificial intelligence training operation includes: Determine the corresponding relationship information between the graphics processor GPU and the host node according to the usage device information in the collective communication log; Determine the source GPU, the destination GPU, the access mode, the access direction, and the access network card according to the communication path information in the collective communication log; Determine, according to the corresponding relationship information, a source host node corresponding to the source GPU and a destination host node corresponding to the destination GPU; Query the basic information of the access network card; The network demand report is generated according to the source GPU, the source host node, the destination GPU, the destination host node, the access mode, the access direction, the access network card and basic information of the access network card.

2. The method according to claim 1, characterized in that The detecting the network status of the network according to the network demand report to obtain the detection result includes: Detect whether each access network card in the network demand report can communicate normally; If it is detected that at least one access network card in the network demand report cannot communicate normally, the access network card that cannot communicate normally is recorded in the detection result; Optimizing the topology of the network according to the detection result includes: If the detection result includes an access network card that cannot communicate normally, the access network card that cannot communicate normally is deleted from the topology structure.

3. The method according to claim 1, characterized in that The detecting the network status of the network according to the network demand report to obtain the detection result includes: Detect whether the network parameters of each access network card in the network demand report are abnormal; If it is detected that the network parameters of at least one access network card in the network demand report are abnormal, the access network card with the abnormal network parameters is recorded in the detection result; The optimizing the communication path of the collective communication according to the state includes: If the detection result includes an access network card with abnormal network parameters, the access network card with abnormal network parameters is deleted from the topology structure.

4. The method according to claim 2, characterized in that: The detecting whether each access network card in the network demand report can communicate normally includes: For each access network card, searching the IP address of the outgoing interface in the routing table of the access network card; determining whether the IP address of the outgoing interface and the IP address of the corresponding gateway are in the same subnet; if the IP address of the outgoing interface and the IP address of the corresponding gateway are not in the same subnet, determining that the access network card cannot communicate normally; and / or For each access network card, it is detected whether the message sent by the access network card can reach the opposite network card; if the message sent by the access network card cannot reach the opposite network card, it is determined that the access network card cannot communicate normally.

5. The method according to claim 3, characterized in that: The detecting whether the network parameters of each access network card in the network demand report are abnormal includes: For each access network card, detecting the bandwidth and / or delay between the access network card and the peer network card; If the bandwidth and / or delay is within an abnormal range, it is determined that the network parameters of the access network card are abnormal.

6. An artificial intelligence computing device, characterized in that: The device comprises: A report generation module, used to generate a network demand report based on the collective communication log corresponding to the artificial intelligence calculation, wherein the network demand report records the demand of the collective communication for the network; A network detection module, used to detect the network status of the network according to the network demand report and obtain a detection result; An optimization module, used for optimizing the topology of the network according to the detection result; A computing module, used to perform artificial intelligence computing based on the optimized network; The report generation module comprises: A relationship determination submodule, used to determine the corresponding relationship information between the graphics processor GPU and the host node according to the usage device information in the collective communication log; A demand determination submodule, used to determine the source GPU, the destination GPU, the access mode, the access direction and the access network card according to the communication path information in the collective communication log; A host determination submodule, used to determine a source host node corresponding to the source GPU and a destination host node corresponding to the destination GPU according to the corresponding relationship information; A query submodule, used to query the basic information of the access network card; A generation submodule is used to generate the network demand report according to the source GPU, the source host node, the destination GPU, the destination host node, the access mode, the access direction, the access network card and basic information of the access network card.

7. An electronic device, characterized in that: The electronic device comprises: processor; a memory for storing processor-executable instructions; Wherein, the processor is configured to execute the artificial intelligence computing method described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a computer to execute the artificial intelligence computing method according to any one of claims 1 to 5.

Images