A multi-GPU BOX management method, product, device and storage medium

Abstract

The present application discloses a multi-GPU BOX management method, product, device and storage medium, which relates to the field of computer technology, including: obtaining identity information sent by each second communication port of multiple GPU BOXs connected to multiple first communication ports on a dual-node server panel; determining the total number of GPU BOXs and the connection mode of each GPU BOX with the first communication port based on the identity information; determining the connection type between the current dual-node server and multiple GPU BOXs based on the total number and the connection mode, and determining the target communication port from multiple first communication ports based on the connection type; switching the target bus link corresponding to the target communication port to the target communication port through a selector and / or a bus arbitrator, so as to perform corresponding channel management on the corresponding GPU BOX through the target communication port. The present application realizes the management of multiple GPU BOXs by a dual-node server.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a multi-GPU BOX management method, product, device and storage medium. Background Art

[0002] With the rapid development of AI (Artificial Intelligence) computing technology, the scale of AI servers that need to be deployed is also increasing, among which GPU BOX with higher density is the main one. GPU BOX has a large number of GPU (Graphic Processing Unit) cards installed inside, which can perform model training, reasoning and other services. In addition, GPU BOX is connected to the server host to realize management, business interaction and other functions.

[0003] In terms of physical implementation, GPU BOX leads out CDFP port (Compact Data Center Fabric Port, a high-speed data transmission port) and connects to CDFP port of server host through DAC (Direct Attach Cable). In addition, in addition to data services, BMC (Baseboard Management Controller) of server host and BMC inside GPU BOX can also perform out-of-band management and communication, and IPMB (Intelligent Platform Management Bus) channel is usually used between BMC of server host and BMC of GPU BOX, and I2C bus (Inter-Integrated Circuit bus) is used physically.

[0004] At present, in a single-node server, the management of GPU BOX can be realized by adding an I2C link to the CDFP port connected to the GPU BOX and connecting it to the BMC of the server motherboard. That is, a single GPU BOX only needs one IPMB channel to realize the host management of GPU BOX. However, in order to improve the density of server hosts in a single cabinet, two server motherboards will be placed in the same server cabinet, that is, a dual-node server. There is currently no combination of connecting GPU BOX for dual-node servers. Dual-node servers have their own hardware and operating systems, which are independent of each other. Therefore, in order to realize the management of multiple GPU BOX by dual-node servers, it is necessary to meet a variety of different adaptation scenarios, such as one server with two GPU BOXs, or one server with four GPU BOXs. In addition, there may be multiple situations for the interconnection between the dual-node server and the GPU BOX. For example, the business link of a GPU BOX is connected to two server nodes at the same time or only to one server node, and its management logic and path are different.

[0005] In summary, how to solve the problem of being unable to judge the current configuration and interconnection method when a dual-node server host is equipped with GPU BOXs with different port numbers and different interconnection methods between the host and the GPU BOX, and thus being unable to manage multiple GPU BOXs is a problem that needs to be further solved in this field. Summary of the invention

[0006] In view of this, the purpose of this application is to provide a multi-GPU BOX management method, product, device and storage medium, which can solve the problem that when a dual-node server host is equipped with GPU BOXs with different port numbers and different interconnection methods between the host and the GPU BOX, the dual-node server cannot determine the current configuration and interconnection method, and thus cannot manage the multi-GPU BOX, thereby realizing the management of the multi-GPU BOX by the dual-node server. The specific scheme is as follows:

[0007] In a first aspect, the present application discloses a multi-GPU BOX management method, which is applied to a dual-node server, comprising:

[0008] Respectively obtain identity information sent by each second communication port of a plurality of GPU BOXs connected to a plurality of first communication ports on the dual-node server panel; the identity information is used to uniquely identify the GPU BOX corresponding to the second communication port;

[0009] Determine the total number of the GPU BOXs and the connection mode between each of the GPU BOXs and the first communication port based on the plurality of identity identification information;

[0010] Determine a connection type between the current dual-node server and the plurality of GPU BOXes based on the total number and the connection mode, and determine a target communication port from the plurality of first communication ports based on the connection type;

[0011] The target bus link corresponding to the target communication port is switched to the target communication port through a selector and / or a bus arbiter, so as to perform corresponding channel management on the corresponding GPU BOX through the target communication port.

[0012] Optionally, the connection type includes a first connection type, a second connection type, a third connection type and a fourth connection type;

[0013] Among them, the first connection type is that a dual-node server is connected to two GPU BOXs, and the communication port of a single GPU BOX is only connected to the communication port of a single server node; the second connection type is that a dual-node server is connected to two GPU BOXs, and the communication port of a single GPU BOX is connected to the communication ports of both server nodes; the third connection type is that a dual-node server is connected to four GPU BOXs, and the communication port of a single GPU BOX is only connected to the communication port of a single server node; the fourth connection type is that a dual-node server is connected to four GPU BOXs, and the communication port of a single GPU BOX is connected to the communication ports of both server nodes.

[0014] Before respectively acquiring the identity information sent by each second communication port of a plurality of GPUBOXs connected to a plurality of first communication ports on the dual-node server panel, the method further includes:

[0015] According to a preset rule, the target communication ports for channel management corresponding to the first connection type, the second connection type, the third connection type and the fourth connection type are respectively selected from the plurality of first communication ports on the dual-node server panel;

[0016] Among them, the number of target communication ports corresponding to the first connection type and the second connection type is 2, the number of target communication ports corresponding to the third connection type and the fourth connection type is 4, and the management channel of each target communication port corresponding to the first connection type and the third connection type is derived from a uniquely determined server node, and the server node from which the management channel of each target communication port corresponding to the second connection type and the fourth connection type comes is not unique; the server node is the first server node or the second server node in the dual-node server.

[0017] Optionally, the determining the connection type between the current dual-node server and the plurality of GPU BOXs based on the total number and the connection mode includes:

[0018] When the total number is 2, and the connection mode is that the multiple second communication ports in each GPU BOX are only connected to the communication port of the first server node or the second server node, it is determined that the current connection type between the dual-node server and the multiple GPU BOXs is the first connection type.

[0019] Optionally, the determining the connection type between the current dual-node server and the plurality of GPU BOXs based on the total number and the connection mode includes:

[0020] When the total number is 2, and the connection mode is that the second communication port in each of the GPU BOXes is connected to the communication ports of the first server node and the second server node, it is determined that the current connection type between the dual-node server and the plurality of GPU BOXes is the second connection type.

[0021] Optionally, the determining the connection type between the current dual-node server and the plurality of GPU BOXs based on the total number and the connection mode includes:

[0022] When the total number is 4, and the connection mode is that the second communication port in each GPU BOX is connected only to the communication port of the first server node or the second server node, it is determined that the current connection type between the dual-node server and the plurality of GPU BOXs is the third connection type.

[0023] Optionally, the determining the connection type between the current dual-node server and the plurality of GPU BOXs based on the total number and the connection mode includes:

[0024] When the total number is 4, and the connection mode is that the second communication port in each of the GPU BOXes is connected to the communication ports of the first server node and the second server node, it is determined that the current connection type between the dual-node server and the plurality of GPU BOXes is the fourth connection type.

[0025] Optionally, switching the target bus link corresponding to the target communication port to the target communication port through a gate includes:

[0026] The corresponding target bus link is enabled, and the target bus link corresponding to the target communication port is switched to the target communication port through a selector.

[0027] Optionally, the switching of the target bus link corresponding to the target communication port to the target communication port through a gate and / or a bus arbiter includes:

[0028] The corresponding target bus link is enabled, and the target bus link corresponding to the target communication port is switched to the target communication port through a selector and / or a bus arbiter.

[0029] Optionally, the multi-GPU BOX management method further includes:

[0030] Detecting whether the data field on the target bus link has been transmitted;

[0031] If the data field on the target bus link has been transmitted, the target bus link is released, and the target bus link is seized by another server node;

[0032] The target bus link is led to the target communication port through the other server node, so as to perform channel management on the corresponding GPU BOX through the other server node and the target communication port.

[0033] Optionally, the multi-GPU BOX management method further includes:

[0034] A preset number of signal pins are reserved in the second communication port of each GPU BOX.

[0035] Optionally, the step of respectively acquiring identity information sent by each second communication port in a plurality of GPU BOXs connected to a plurality of first communication ports on the dual-node server panel includes:

[0036] The identity information generated and sent by each second communication port in a plurality of GPU BOXs connected to a plurality of first communication ports on the dual-node server panel after performing resistor pull-up and pull-down processing on a preset number of signal pins of the second communication port is respectively obtained.

[0037] Optionally, the determining the total number of the GPU BOXs and the connection mode between each GPU BOX and the first communication port based on the plurality of identity identification information includes:

[0038] Sorting the plurality of identity identification information corresponding to the GPU BOX based on the order of the plurality of first communication ports to obtain sorted identification information;

[0039] The total number of the plurality of GPU BOXs and the connection mode between each of the GPU BOXs and the first communication port are determined based on the sorted identification information.

[0040] In a second aspect, the present application discloses a computer program product, including a computer program / instruction, which implements the aforementioned multi-GPU BOX management method when executed by a processor.

[0041] In a third aspect, the present application discloses an electronic device, including a processor and a memory; wherein the processor implements the aforementioned multi-GPU BOX management method when executing a computer program stored in the memory.

[0042] In a fourth aspect, the present application discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, the aforementioned multi-GPU BOX management method is implemented.

[0043] It can be seen that the present application is applied to a dual-node server, firstly, the identity identification information sent by each second communication port of a plurality of GPU BOXs connected to a plurality of first communication ports on a panel of the dual-node server is respectively obtained, and then the total number of the GPU BOXs and the connection mode of each GPU BOX and the first communication port are determined based on the plurality of identity identification information, and the connection type between the current dual-node server and the plurality of GPU BOXs is determined based on the total number and the connection mode, and the target communication port is determined from the plurality of first communication ports based on the connection type, and then the target bus link corresponding to the target communication port is switched to the target communication port through a selector and / or a bus arbiter, so as to perform corresponding channel management on the corresponding GPU BOX through the target communication port. When implementing channel management of multiple GPU BOXes by a dual-node server, the present application determines the total number of GPU BOXes and the connection mode of each GPU BOX with the communication port on the server panel according to the identity identification information sent by each communication port in the multiple GPU BOXes, and determines the connection type between the dual-node server and the multiple GPU BOXes based on the total number and the connection mode. Finally, the target bus link corresponding to the connection type is switched to the corresponding target communication port through a selector and / or a bus arbitrator. Through the above method, the problem that the dual-node server cannot judge the current configuration and interconnection method, and thus cannot manage the multiple GPU BOXes when the dual-node server host is equipped with GPUBOXs of different port numbers and different interconnection methods between the host and the GPU BOXes can be solved, thereby realizing the management of multiple GPU BOXes by the dual-node server, and being applicable to various application scenarios with different numbers of GPU BOXes, different numbers of GPU BOX interfaces, and different GPU BOX interconnection methods, so that the dual-node server can be equipped with GPU BOXes of any form, any number, and any connection method, thereby filling the gap in the field of dual-node server with multiple GPU BOXes. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0045] Figure 1 A flowchart of a multi-GPU BOX management method disclosed in this application;

[0046] Figure 2 A schematic diagram of a connection method of a dual-node server internal node to a server panel port disclosed in the present application, and different connection methods of a dual-node server to a GPU BOX;

[0047] Figure 3 A schematic diagram of a first connection type of connection between a dual-node server and two GPU BOXes disclosed in the present application;

[0048] Figure 4 A schematic diagram of a second connection type of connection between a dual-node server and two GPU BOXes disclosed in the present application;

[0049] Figure 5 A schematic diagram of a third connection type of connection between a dual-node server and four GPU BOXes disclosed in the present application;

[0050] Figure 6 A schematic diagram of a fourth connection type of connection between a dual-node server and four GPU BOXes disclosed in the present application;

[0051] Figure 7 A schematic diagram of a GPU BOX ID identified in a different connection mode disclosed in this application;

[0052] Figure 8 A specific multi-GPU BOX management flow chart disclosed in this application;

[0053] Fig. 9 Disclosed for this application Figure 8 Schematic diagram of the functional annotations of each connection and module;

[0054] Fig.10 A flowchart of a multi-GPU BOX management method disclosed in this application;

[0055] Fig.11 This is a structural diagram of an electronic device disclosed in this application. DETAILED DESCRIPTION

[0056] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0057] The present application embodiment discloses a multi-GPU BOX management method applied to a dual-node server, see Figure 1 As shown, the method includes:

[0058] Step S11: respectively obtaining identity information sent by each second communication port of a plurality of GPU BOXs connected to a plurality of first communication ports on the dual-node server panel; the identity information is used to uniquely identify the GPU BOX corresponding to the second communication port.

[0059] It should be noted that the multi-GPU BOX management solution proposed in this application is applied to a dual-node server, which includes two server nodes, namely a first server node and a second server node, and the dual-node server panel includes multiple communication ports (such as CDFP ports), which can be connected to the communication ports (such as CDFP ports) on the GPU BOX through a bus (such as an I2C bus) to perform corresponding channel management, such as IPMB channel management. For details, see Figure 2 As shown, the dual-node server includes two server nodes, namely server node A and server node B, and the panel of the dual-node server includes 8 CDFP ports, wherein the data link inside server node A corresponds to port 1, port 2, port 3, and port 4, and the data link inside server node B corresponds to port 5, port 6, port 7, and port 8. In addition, Figure 2 Two different connection modes between GPU BOX and dual-node servers are shown. Connection mode A is that all CDFP ports of GPU BOX are connected to the ports of the data link corresponding to server node A, such as port 1 and port 2 of GPU BOX are connected to port 1 and port 2 of server node A respectively; connection mode B is that all CDFP ports of GPU BOX are connected to the ports of the data links corresponding to server node A and server node B, such as port 1 of GPU BOX is connected to port 1 of server node A, and port 2 of GPU BOX is connected to port 5 of server node B at the same time.

[0060] In this embodiment, the communication port in each GPU BOX will send identity information to the corresponding communication port on the dual-node server panel. The identity information is used to uniquely identify the GPUBOX corresponding to the communication port that currently sends the information, that is, the ID (Identity document) of the GPU BOX. Therefore, the dual-node server can receive the identity information sent by each second communication port in the multiple GPU BOXs connected to the multiple first communication ports on the current dual-node server panel. It should be pointed out that the identity information sent by the multiple communication ports from the same GPU BOX is the same, and all identify the same GPU BOX.

[0061] Specifically, before performing multi-GPU BOX management, the method may further include: reserving a preset number of signal pins in the second communication port of each GPU BOX, that is, reserving a preset number of signal pins in each communication port of the GPU BOX in advance.

[0062] Correspondingly, the step of respectively obtaining the identity information sent by each second communication port of the plurality of GPU BOXs connected to the plurality of first communication ports on the dual-node server panel may specifically include: respectively obtaining the identity information generated and sent by each second communication port of the plurality of GPU BOXs connected to the plurality of first communication ports on the dual-node server panel after performing a resistor pull-up / down process on a preset number of signal pins of the second communication port. In this embodiment, a corresponding identity code may be set in advance for each GPU BOX as an identifier to distinguish it from other GPU BOXs. When a preset time period is reached, three preset signal pins are selected from the signal pins of the CDFP port of the GPU BOX, and the ID of the corresponding GPU BOX is generated by performing a resistor pull-up / down process on the three signal pins. For example, the level state of the signal pin is changed by the resistor pull-up / down process to generate a level signal of 010, and the level signal (i.e., 010) is used as the ID of the corresponding GPU BOX. In this way, when the communication port on the dual-node server panel receives the above-mentioned level signal sent by each communication port in the GPU BOX, it can be known which GPU BOX the signal comes from, the number of ports of each GPU BOX, and the connection mode of the communication ports between the two.

[0063] Step S12: determining the total number of the GPU BOXes and the connection mode between each GPU BOX and the first communication port based on the plurality of identity identification information.

[0064] In this embodiment, after obtaining the identity identification information sent by each second communication port in the multiple GPU BOXs, all the received identity identification information can be sorted first, and the sorting rule can be based on the arrangement order of the multiple communication ports on the dual-node server panel, such as first sorting according to the order of the communication ports of the first server node in the dual-node server, and then sorting according to the order of the communication ports of the second server node in the dual-node server, and then determining the total number of all GPU BOXs and the connection mode between each GPU BOX and the multiple first communication ports on the dual-node server panel according to the sorted identity identification information. For example, after receiving the identity information sent by each communication port in multiple GPU BOXes, they are sorted according to the order of each communication port on the dual-node server panel to obtain 001, 010, 011, 100, 001, 010, 011, 100. Through analysis, it can be known that the total number of current GPU BOXes is 4 (i.e., the number of different identity information), the total number of communication ports on the dual-node server panel is 8 (i.e., the number of all identity information), and the number of communication ports of each GPU BOX is 2, and the communication ports corresponding to the two server nodes are connected (where 001, 010, 011, 100 correspond to the first server node in the dual-node server, and 001, 010, 011, 100 correspond to the second server node in the dual-node server).

[0065] Step S13: determining the connection type between the current dual-node server and the plurality of GPU BOXs based on the total number and the connection mode, and determining a target communication port from the plurality of first communication ports based on the connection type.

[0066] In this embodiment, after the total number of the GPU BOXs and the connection mode between each of the GPU BOXs and the first communication port are determined, the connection type between the current dual-node server and the multiple GPU BOXs is further determined based on the total number and the connection mode, and the target communication port for channel management is determined from the multiple first communication ports on the dual-node server panel based on the connection type.

[0067] The connection types specifically include a first connection type, a second connection type, a third connection type and a fourth connection type; specifically, the first connection type is that a dual-node server is connected to two GPU BOXs, and the communication port of each GPUBOX is only connected to the communication port of one server node, for example, see Figure 3As shown, a dual-node server (including server node A and server node B) is connected to two GPU BOXs, namely GPU BOX1 and GPU BOX2, and each GPUBOX includes four communication ports, wherein the four communication ports of GPU BOX1 are connected to the four communication ports corresponding to server node A (i.e., ports 1, 2, 3, and 4), and the four communication ports of GPU BOX2 are connected to the four communication ports corresponding to server node B (i.e., ports 5, 6, 7, and 8).

[0068] Specifically, the second connection type is that a dual-node server is connected to two GPU BOXes, and the communication port of a single GPUBOX is connected to the communication ports of the two server nodes. For example, see Figure 4 As shown, a dual-node server (including server node A and server node B) is connected to two GPU BOXs, namely GPU BOX1 and GPU BOX2, and each GPUBOX includes four communication ports, wherein port 1 and port 2 in GPU BOX1 are respectively connected to port 1 and port 2 of server node A, and port 3 and port 4 in GPU BOX1 are respectively connected to port 5 and port 6 of server node B, that is, it is connected to the communication ports of the two server nodes at the same time; in addition, GPU BOX2 is connected in the same way as GPU BOX1, and is connected to the communication ports of the two server nodes at the same time, such as port 1 and port 2 in GPU BOX2 are respectively connected to port 3 and port 4 of server node A, and port 3 and port 4 in GPU BOX1 are respectively connected to port 7 and port 8 of server node B.

[0069] Specifically, the third connection type is that a dual-node server is connected to four GPU BOXes, and the communication port of a single GPUBOX is only connected to the communication port of a single server node, for example, see Figure 5 As shown, a dual-node server (including server node A and server node B) is connected to four GPU BOXs, namely GPU BOX1, GPU BOX2, GPU BOX3 and GPU BOX4, and each GPU BOX includes two communication ports, wherein port 1 and port 2 in GPU BOX1 are respectively connected to port 1 and port 2 of server node A, port 1 and port 2 in GPU BOX2 are respectively connected to port 5 and port 6 of server node B, port 1 and port 2 in GPU BOX3 are respectively connected to port 3 and port 4 of server node A, and port 1 and port 2 in GPUBOX4 are respectively connected to port 7 and port 8 of server node B.

[0070] Specifically, the fourth connection type is that a dual-node server is connected to four GPU BOXes, and the communication port of a single GPUBOX is connected to the communication ports of two server nodes, for example, see Figure 6 As shown, a dual-node server (including server node A and server node B) is connected to four GPU BOXs, namely GPU BOX1, GPU BOX2, GPU BOX3 and GPU BOX4, and each GPU BOX includes two communication ports, wherein port 1 and port 2 in GPU BOX1 are respectively connected to port 1 of server node A and port 5 of server node B, port 1 and port 2 in GPU BOX2 are respectively connected to port 2 of server node A and port 6 of server node B, port 1 and port 2 in GPU BOX3 are respectively connected to port 3 of server node A and port 7 of server node B, and port 1 and port 2 in GPU BOX4 are respectively connected to port 4 of server node A and port 8 of server node B.

[0071] In this embodiment, by analyzing and classifying the dual-node server host with GPU BOXes of different port numbers and different interconnection methods between the host (i.e., server node) and the GPU BOX, the dual-node server can manage multiple GPU BOXes. This is applicable to various application scenarios with different numbers of GPU BOXes, different numbers of GPU BOX interfaces, and different GPUBOX interconnection methods, and improves the efficiency of multiple GPU BOX management.

[0072] It is understandable that for different connection methods, the corresponding identity information is also different, see Figure 7 As shown in the figure, when the received identity information is sorted in the order of the interfaces as 001, 001, 001, 001, 010, 010, 010, 010, it can be known that the total number of the current GPU BOX is 2, the total number of communication ports on the dual-node server panel is 8, and the number of communication ports of each GPUBOX is 4, and they are connected to the 4 communication ports corresponding to a single server node. Therefore, it can be determined that the current connection type is the first connection type (corresponding to Figure 3 Similarly, when the received identity information is sorted according to the interface sequence as 001, 001, 010, 010, 001, 001, 010, 010, it can be known that the total number of current GPU BOXs is 2, the total number of communication ports on the dual-node server panel is 8, and the number of communication ports of each GPU BOX is 4, and the communication ports corresponding to the two server nodes are connected, so it can be determined that the current connection type is the second connection type (corresponding to Figure 4Similarly, when the received identity information is sorted in the order of the interfaces as 001, 001, 011, 011, 010, 010, 100, 100, it can be known that the total number of current GPU BOXs is 4, the total number of communication ports on the dual-node server panel is 8, and the number of communication ports of each GPU BOX is 2, and they are connected to the 2 communication ports corresponding to a single server node. Therefore, it can be determined that the current connection type is the third connection type (corresponding to Figure 5 Similarly, when the received identity information is sorted in the order of the interfaces as 001, 010, 011, 100, 001, 010, 011, 100, it can be known that the total number of GPU BOXs is 4, the total number of communication ports on the dual-node server panel is 8, and the number of communication ports of each GPU BOX is 2, and the communication ports corresponding to the two server nodes are connected, so it can be determined that the current connection type is the fourth connection type (corresponding to Figure 6 connection method).

[0073] In addition, before respectively acquiring the identity information sent by each second communication port in the plurality of GPU BOXs connected to the plurality of first communication ports on the dual-node server panel, the method further includes: filtering out the target communication ports for channel management corresponding to the first connection type, the second connection type, the third connection type and the fourth connection type from the plurality of first communication ports on the dual-node server panel according to a preset rule; wherein the number of the target communication ports corresponding to the first connection type and the second connection type is 2, the number of the target communication ports corresponding to the third connection type and the fourth connection type is 4, and the management channel of each of the target communication ports corresponding to the first connection type and the third connection type is derived from a uniquely determined server node, and the server node from which the management channel of each of the target communication ports corresponding to the second connection type and the fourth connection type is derived is not unique; the server node is the first server node in the dual-node server (i.e. Figure 2 Server node A in the Figure 2 It should be noted that the present application sets target communication ports for channel management for server nodes and GPU BOX corresponding to different connection types according to the preset rules, wherein the number of target communication ports corresponding to the first connection type and the second connection type is 2, and the number of target communication ports corresponding to the third connection type and the fourth connection type is 4. For example, see Figure 3As shown, when it is the first connection type, according to the preset rules, the path between CDFP port 1 of GPU BOX1 and CDFP port 1 of server node A is used as the IPMB management path, and CDFP port 1 of server node A is set as the target communication port (that is, the frontmost port in a group of CDFP ports connected to each GPU BOX is selected as the management channel interface). Similarly, CDFP port 5 of server node B is set as the target communication port, and the number of target communication ports corresponding to the first connection type is 2; when it is the second connection type, see Figure 4 As shown, CDFP ports 1 and 3 of server node A are set as target communication ports, then the number of target communication ports corresponding to the second connection type is 2; when it is the third connection type, see Figure 5 As shown, CDFP ports 1 and 3 of server node A and CDFP ports 5 and 7 of server node B are set as target communication ports, then the number of target communication ports corresponding to the third connection type is 4; when it is the fourth connection type, see Figure 6 As shown, CDFP ports 1, 2, 3, and 4 of server node A are set as target communication ports, and the number of target communication ports corresponding to the fourth connection type is 4.

[0074] In addition, it should be noted that the management channels of the target communication ports corresponding to the first connection type and the third connection type are derived from a uniquely determined server node. Figure 3 and Figure 5 As shown, Figure 3 The IPMB management channel of CDFP port 1 on the server panel is derived from the uniquely determined server node A, and the IPMB management channel of CDFP port 5 on the server panel is derived from the uniquely determined server node B; or Figure 5 The IPMB management channels of CDFP ports 1 and 3 on the server panel are derived from the uniquely determined server node A, and CDFP ports 5 and 7 on the server panel are derived from the uniquely determined server node B.

[0075] Furthermore, the server nodes from which the management channels of the target communication ports corresponding to the second connection type and the fourth connection type come are not unique (they may come from server node A or server node B). Figure 4 and Figure 6 As shown, the management channel of the target communication port corresponding to the third connection type and the fourth connection type may come from the server node A or the server node B, such as Figure 4The IPMB management paths of CDFP ports 1 and 3 on the server panel may come from server node A or server node B. The specific path to be selected needs to be determined by corresponding arbitration and / or gating during the real-time multi-GPU BOX management process. Figure 6 The IPMB management paths of CDFP ports 1, 2, 3, and 4 on the server panel may originate from server node A or server node B.

[0076] In a first specific implementation manner, the determining of the connection type between the current dual-node server and the multiple GPU BOXes based on the total number and the connection mode may specifically include: when the total number is 2, and the connection mode is that the multiple second communication ports in each GPU BOX are only connected to the communication port of the first server node or the second server node, determining that the connection type between the current dual-node server and the multiple GPU BOXes is the first connection type. For example, see Figure 3 As shown, when the number of all GPU BOXs is 2, and the connection mode is that the 4 CDFP communication ports in a single GPU BOX are only connected to the CDFP communication port of server node A or server node B, it is determined that the connection type between the current dual-node server and the multi-GPU BOX is the first connection type.

[0077] In a second specific implementation manner, the determining of the connection type between the current dual-node server and the multiple GPU BOXes based on the total number and the connection mode may specifically include: when the total number is 2, and the connection mode is that the second communication port in each of the GPU BOXes is connected to the communication ports of the first server node and the second server node, determining that the connection type between the current dual-node server and the multiple GPU BOXes is the second connection type. For example, see Figure 4 As shown, when the number of all GPU BOXes is 2, and the connection mode is that the 4 CDFP communication ports in a single GPU BOX are connected to the CDFP communication ports of server node A and server node B, it is determined that the connection type between the current dual-node server and the multi-GPU BOX is the second connection type.

[0078] In a third specific implementation manner, the determining of the connection type between the current dual-node server and the multiple GPU BOXes based on the total number and the connection mode may specifically include: when the total number is 4, and the connection mode is that the second communication port in each GPU BOX is only connected to the communication port of the first server node or the second server node, determining that the connection type between the current dual-node server and the multiple GPU BOXes is the third connection type. For example, see Figure 5 As shown, when the number of all GPU BOXes is 4, and the connection mode is that 2 CDFP communication ports in a single GPU BOX are connected to the CDFP communication port of server node A or server node B, it is determined that the connection type between the current dual-node server and the multi-GPU BOX is the third connection type.

[0079] In a fourth specific implementation manner, the determining of the connection type between the current dual-node server and the multiple GPU BOXes based on the total number and the connection mode may specifically include: when the total number is 4, and the connection mode is that the second communication port in each of the GPU BOXes is connected to the communication ports of the first server node and the second server node, determining that the connection type between the current dual-node server and the multiple GPU BOXes is the fourth connection type. For example, see Figure 6 As shown, when the number of all GPU BOXes is 4, and the connection mode is that the 2 CDFP communication ports in a single GPU BOX are connected to the CDFP communication ports of server node A and server node B, it is determined that the connection type between the current dual-node server and the multi-GPU BOX is the fourth connection type.

[0080] Step S14: switching the target bus link corresponding to the target communication port to the target communication port through a selector and / or a bus arbiter, so as to perform corresponding channel management on the corresponding GPU BOX through the target communication port.

[0081] In this embodiment, after determining a target communication port from the plurality of first communication ports based on the connection type, the target bus link corresponding to the target communication port can be switched to the target communication port through a selector and / or a bus arbiter, thereby performing corresponding channel management on the corresponding GPU BOX through the target communication port.

[0082] In this embodiment, when the connection type is the first connection type, switching the target bus link corresponding to the target communication port to the target communication port through the gate may specifically include: enabling the corresponding target bus link, and switching the target bus link corresponding to the target communication port to the target communication port through the gate. Figure 8 As shown, for the first connection type, the ID sequence of the GPU BOX can be sorted by the gate to obtain the true value, and then the current connection mode is identified based on the sorted true value and the bus link is automatically switched, as shown in FIG. Figure 8 After determining the target communication port (i.e. CDFP port 1 and 5 on the server panel), first enable the target bus link (i.e. I2C_A1 and I2C_B1) of the corresponding CDFP port, and then switch the target bus link (i.e. I2C_A1 and I2C_B1) corresponding to the target communication port to the target communication port (i.e. CDFP port 1 and 5 on the server panel) through the strobe switch (i.e. gate). In addition, see Fig. 9 As shown, Figure 8 The dotted line in the middle represents the I2C link, the solid line represents the I2C link that is ultimately connected to the CDFP port, the dotted box represents that the CDFP port is used as an IPMB management port under a certain connection method, and the solid box represents a common CDFP data interface, which will not be used as an IPMB management port under any connection method.

[0083] In this embodiment, when the connection type is the third connection type, switching the target bus link corresponding to the target communication port to the target communication port through the gate may specifically include: enabling the corresponding target bus link, and switching the target bus link corresponding to the target communication port to the target communication port through the gate. Figure 8 As shown, for the third connection type, similar to the first connection type, after the target communication port (i.e., CDFP ports 1, 3, 5, and 7 on the server panel) is determined, the target bus links (i.e., I2C_A1, I2C_A3, I2C_B1, and I2C_B3) of the corresponding CDFP ports are first enabled, and then the target bus links (i.e., I2C_A1, I2C_A3, I2C_B1, and I2C_B3) corresponding to the target communication ports are switched to the target communication ports (i.e., CDFP ports 1, 3, 5, and 7 on the server panel) through a selection switch (selector).

[0084] In this embodiment, when the connection type is the second connection type, switching the target bus link corresponding to the target communication port to the target communication port through the gate and / or the bus arbiter may specifically include: enabling the corresponding target bus link, and switching the target bus link corresponding to the target communication port to the target communication port through the gate and / or the bus arbiter. Figure 8 As shown, for the second connection type, after determining the target communication port (i.e., CDFP ports 1 and 3 on the server panel), first enable the target bus link (i.e., I2C_A1 and I2C_A3) of the corresponding CDFP port, then sort the ID sequence of the GPU BOX through the selector to obtain a true value, and identify the current connection mode based on the sorted true value, and then lead the I2C link to the corresponding arbiter through the selector to arbitrate the bus source corresponding to the target communication port through the arbiter, and then lead the target I2C link (i.e., I2C_A1 or I2C_A3) corresponding to the bus source to the target communication port (i.e., CDFP ports 1 and 3 on the server panel).

[0085] In this embodiment, when the connection type is the fourth connection type, switching the target bus link corresponding to the target communication port to the target communication port through the gate and / or the bus arbiter may specifically include: enabling the corresponding target bus link, and switching the target bus link corresponding to the target communication port to the target communication port through the gate and / or the bus arbiter. Figure 8 As shown, for the fourth connection type, after determining the target communication port (i.e., CDFP ports 1, 2, 3, and 4 on the server panel), first enable the target bus link (i.e., I2C_A1, I2C_A2, I2C_A3, and I2C_A4) of the corresponding CDFP port, and then enable the GPU through the gating switch (gating device). The ID sequence of the BOX is sorted to obtain the truth value, and the current connection mode is identified based on the sorted truth value. Then, the I2C link (i.e., I2C_A1 and I2C_A3) is led to the corresponding arbiter through the selector, so as to arbitrate the bus source corresponding to the target communication port through the arbitrator, and then the target I2C link (i.e., I2C_A1 or I2C_A3) corresponding to the bus source is led to the target communication port (i.e., CDFP port 1 or 3 on the server panel); at the same time, the I2C link (i.e., I2C_A2 and I2C_A4) is arbitrated through the arbitrator, and based on the bus source corresponding to the arbitrated target communication port, the corresponding target I2C link (i.e., I2C_A2 or I2C_A4) is led to the target communication port (i.e., CDFP port 2 or 4 on the server panel).

[0086] In addition, it should be pointed out that when data transmission is performed on the target bus link, the following may also be included: detecting whether the data field on the target bus link has been transmitted; if the data field on the target bus link has been transmitted, releasing the target bus link, and seizing the target bus link through another server node; and guiding the target bus link to the target communication port through the other server node, so as to perform channel management on the corresponding GPU BOX through the other server node and the target communication port. For example, the bus arbiter in the dual-node server detects whether the data field on the target bus link (transmitted by server node A) has been transmitted. If the data field on the target bus link has been transmitted, releasing the target bus link (such as the bus after I2C_A1 passes through the gate), seizing the target bus link through server node B, and then guiding the target bus link (such as the bus after I2C_B1 passes through the gate) to the target communication port (such as CDFP port 1 on the server panel) through server node B, so as to manage the corresponding GPU BOX (such as Figure 6 It should be noted that the bus arbiter allows node A to obtain management rights by default, that is, the I2C link of server node A is selected as the final output and led to the CDFP port of the server. However, the preemption logic can also be implemented. After the data field on the I2C link of server node A is transmitted, the corresponding I2C link is released. At this time, server node B can preempt the I2C link, and then lead the I2C link of server node B to the CDFP port of the server, so as to realize the management of GPU BOX by server node B.

[0087] It can be seen that the embodiment of the present application is applied to a dual-node server, firstly, the identity identification information sent by each second communication port of a plurality of GPU BOXs connected to a plurality of first communication ports on a panel of the dual-node server is respectively obtained, and then the total number of the GPU BOXs and the connection mode of each GPU BOX and the first communication port are determined based on the plurality of identity identification information, and the connection type between the current dual-node server and the plurality of GPU BOXs is determined based on the total number and the connection mode, and the target communication port is determined from the plurality of first communication ports based on the connection type, and then the target bus link corresponding to the target communication port is switched to the target communication port through a selector and / or a bus arbiter, so as to perform corresponding channel management on the corresponding GPU BOX through the target communication port. In the embodiment of the present application, when a dual-node server performs channel management on multiple GPU BOXes, the total number of GPU BOXes and the connection mode between each GPU BOX and the communication port on the server panel are determined according to the identity identification information sent by each communication port in the multiple GPU BOXes, and the connection type between the dual-node server and the multiple GPU BOXes is determined based on the total number and the connection mode. Finally, the target bus link corresponding to the connection type is switched to the corresponding target communication port through the selector and / or the bus arbiter. Through the above method, the problem that the dual-node server cannot judge the current configuration and interconnection method and thus cannot manage the multiple GPU BOXes when the dual-node server host is equipped with GPU BOXes with different port numbers and different interconnection methods between the host and the GPU BOXes can be solved, thereby realizing the management of the multiple GPU BOXes by the dual-node server, and being applicable to various application scenarios with different numbers of GPU BOXes, different numbers of GPU BOX interfaces, and different GPU BOX interconnection methods, so that the dual-node server can be equipped with GPU BOXes of any form, any number, and any connection method, thereby filling the gap in the field of dual-node server with multiple GPU BOXes.

[0088] The present application embodiment discloses a specific multi-GPU BOX management method applied to a dual-node server, see Fig.10 As shown, the method includes:

[0089] Step S21: respectively obtaining identity information sent by each second communication port of a plurality of GPU BOXs connected to a plurality of first communication ports on the dual-node server panel; the identity information is used to uniquely identify the GPU BOX corresponding to the second communication port.

[0090] Step S22: sorting the plurality of identity identification information corresponding to the GPU BOX based on the order of the plurality of first communication ports to obtain sorted identification information.

[0091] In this embodiment, after obtaining the identity information sent by each second communication port of the multiple GPU BOXes connected to the multiple first communication ports on the dual-node server panel, further, all the identity information corresponding to the above GPU BOXes is sorted based on the order of each first communication port to obtain sorted identification information. Figure 7 As shown, Figure 7 The figure shows the sorted identification information (i.e., sorted GPU BOX ID) under four different connection methods, among which the first connection method is the identification information obtained by sorting all the identity identification information (i.e., GPU BOX ID) corresponding to the GPU BOX in order from left to right and from top to bottom, i.e., 001, 001, 001, 010, 010, 010, 010.

[0092] Step S23: determining the total number of the plurality of GPU BOXs and the connection mode between each GPU BOX and the first communication port based on the sorted identification information.

[0093] Further, based on the above sorted identification information (i.e., 001, 001, 001, 010, 010, 010, 010), it is determined that the total number of all GPU BOXes is 2, the total number of communication ports on the dual-node server panel is 8, and the number of communication ports of each GPU BOX is 4, and they are connected to the 4 communication ports corresponding to a single server node. Therefore, it can be determined that the current connection type is the first connection type (corresponding to Figure 3 Similarly, we can determine Figures 4 to 6 The corresponding connection method.

[0094] Step S24: determining the connection type between the current dual-node server and the plurality of GPU BOXs based on the total number and the connection mode, and determining a target communication port from the plurality of first communication ports based on the connection type.

[0095] Step S25: switching the target bus link corresponding to the target communication port to the target communication port through a selector and / or a bus arbiter, so as to perform corresponding channel management on the corresponding GPU BOX through the target communication port.

[0096] For more specific processing procedures of the above steps S21, S24, and S25, reference may be made to the corresponding contents disclosed in the aforementioned embodiments, which will not be described in detail here.

[0097] It can be seen that the embodiment of the present application first obtains the identity identification information sent by each second communication port of the multiple GPU BOXs connected to the multiple first communication ports on the dual-node server panel respectively, and then sorts the multiple identity identification information corresponding to the GPU BOX based on the order of the multiple first communication ports to obtain sorted identification information, and determines the total number of the multiple GPU BOXs and the connection mode of each GPU BOX and the first communication port based on the sorted identification information, and then determines the connection type between the current dual-node server and the multiple GPU BOXs based on the total number and the connection mode, and determines the target communication port from the multiple first communication ports based on the connection type, and finally switches the target bus link corresponding to the target communication port to the target communication port through the selector and / or the bus arbiter, so as to perform corresponding channel management on the corresponding GPU BOX through the target communication port. The embodiment of the present application sorts the identity information sent by each communication port in the received multiple GPU BOXes, and determines the total number of the multiple GPU BOXes and the connection method between each GPU BOX and the multiple communication ports on the server panel based on the sorted identification information, so as to accurately identify the configuration status and interconnection method of the current dual-node server and the multiple GPUBOXes, thereby realizing the management of the multiple GPU BOXes by the dual-node server.

[0098] Furthermore, the present application also discloses an electronic device. Fig.11 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content in the diagram cannot be regarded as any limitation on the scope of use of the present application.

[0099] Fig.11 The present invention provides a schematic diagram of the structure of an electronic device 20 provided in an embodiment of the present application. The electronic device 20 may specifically include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 is used to store a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the multi-GPU BOX management method disclosed in any of the aforementioned embodiments. In addition, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0100] In this embodiment, the power supply 23 is used to provide working voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and the external device, and the communication protocol it follows is any communication protocol that can be applied to the technical solution of the present application, and is not specifically limited here; the input and output interface 25 is used to obtain external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs and is not specifically limited here.

[0101] In addition, the memory 22, as a carrier for storing resources, can be a read-only memory, a random access memory, a disk or an optical disk, etc. The resources stored thereon can include an operating system 221, a computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0102] The operating system 221 is used to manage and control the hardware devices on the electronic device 20 and the computer program 222, which can be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program that can be used to complete the multi-GPU BOX management method performed by the electronic device 20 disclosed in any of the aforementioned embodiments, the computer program 222 can further include a computer program that can be used to complete other specific tasks.

[0103] Furthermore, the present application also discloses a computer-readable storage medium for storing a computer program; wherein the computer program, when executed by a processor, implements the multi-GPU BOX management method disclosed above. The specific steps of the method can refer to the corresponding contents disclosed in the above embodiments, and will not be repeated here.

[0104] Furthermore, an embodiment of the present application also discloses a computer program product, including a computer program / instruction, which implements the steps of the multi-GPU BOX management method disclosed above when executed by a processor.

[0105] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the products disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method part.

[0106] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0107] The steps of the method or algorithm described in conjunction with the embodiments disclosed herein may be implemented directly using hardware, a software module executed by a processor, or a combination of the two. The software module may be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0108] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0109] The above is a detailed introduction to a multi-GPU BOX management method, product, device and storage medium provided by the present application. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method and core idea of ​​the present application. At the same time, for those skilled in the art, according to the idea of ​​the present application, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as limiting the present application.

Claims

1. A multi-GPU BOX management method, characterized in that: Applicable to dual-node servers, including: Respectively obtain identity information sent by each second communication port of a plurality of GPU BOXs connected to a plurality of first communication ports on the dual-node server panel; the identity information is used to uniquely identify the GPU BOX corresponding to the second communication port; Determine the total number of the GPU BOXs and the connection mode between each of the GPU BOXs and the first communication port based on the plurality of identity identification information; Determine a connection type between the current dual-node server and the plurality of GPU BOXes based on the total number and the connection mode, and determine a target communication port from the plurality of first communication ports based on the connection type; Switching the target bus link corresponding to the target communication port to the target communication port through a selector and / or a bus arbiter, so as to perform corresponding channel management on the corresponding GPU BOX through the target communication port; The connection types include a first connection type, a second connection type, a third connection type and a fourth connection type; wherein the first connection type is that a dual-node server is connected to two GPU BOXs, and the communication port of a single GPU BOX is only connected to the communication port of a single server node; the second connection type is that a dual-node server is connected to two GPU BOXs, and the communication port of the single GPU BOX is connected to the communication ports of both server nodes; the third connection type is that a dual-node server is connected to four GPU BOXs, and the communication port of a single GPU BOX is only connected to the communication port of a single server node; the fourth connection type is that a dual-node server is connected to four GPU BOXs, and the communication port of a single GPU BOX is connected to the communication ports of both server nodes.

2. The multi-GPU BOX management method according to claim 1, characterized in that: Before respectively acquiring the identity information sent by each second communication port of a plurality of GPU BOXs connected to a plurality of first communication ports on the dual-node server panel, the method further includes: According to preset rules, the target communication ports for channel management corresponding to the first connection type, the second connection type, the third connection type and the fourth connection type are respectively selected from the multiple first communication ports on the dual-node server panel; wherein the number of the target communication ports corresponding to the first connection type and the second connection type is 2, the number of the target communication ports corresponding to the third connection type and the fourth connection type is 4, and the management channels of the respective target communication ports corresponding to the first connection type and the third connection type are derived from a uniquely determined server node, while the server nodes from which the management channels of the respective target communication ports corresponding to the second connection type and the fourth connection type are derived are not unique; the server node is the first server node or the second server node in the dual-node server.

3. The multi-GPU BOX management method according to claim 2, characterized in that: When the total number is 2 or 4, determining the connection type between the current dual-node server and the plurality of GPU BOXs based on the total number and the connection mode includes: If the connection mode is that the multiple second communication ports in each GPU BOX are only connected to the communication port of the first server node or the second server node, it is determined that the current connection type between the dual-node server and the multiple GPU BOXs is the first connection type or the third connection type.

4. The multi-GPU BOX management method according to claim 2, characterized in that: When the total number is 2 or 4, determining the connection type between the current dual-node server and the plurality of GPU BOXs based on the total number and the connection mode includes: If the connection mode is that the second communication port in each GPU BOX is connected to the communication ports of the first server node and the second server node, it is determined that the current connection type between the dual-node server and the plurality of GPU BOXs is the second connection type or the fourth connection type.

5. The multi-GPU BOX management method according to claim 3, characterized in that: Switching the target bus link corresponding to the target communication port to the target communication port through a selector includes: The corresponding target bus link is enabled, and the target bus link corresponding to the target communication port is switched to the target communication port through a selector.

6. The multi-GPU BOX management method according to claim 4, characterized in that: The step of switching the target bus link corresponding to the target communication port to the target communication port through a gate and / or a bus arbiter includes: The corresponding target bus link is enabled, and the target bus link corresponding to the target communication port is switched to the target communication port through a selector and / or a bus arbiter.

7. The multi-GPU BOX management method according to claim 6, characterized in that: Also includes: Detecting whether the data field on the target bus link has been transmitted; If the data field on the target bus link has been transmitted, the target bus link is released, and the target bus link is seized by another server node; The target bus link is led to the target communication port through the other server node, so as to perform channel management on the corresponding GPU BOX through the other server node and the target communication port.

8. The multi-GPU BOX management method according to claim 1, characterized in that: Also includes: A preset number of signal pins are reserved in the second communication port of each GPU BOX.

9. The multi-GPU BOX management method according to claim 8, characterized in that: The step of respectively acquiring identity information sent by each second communication port in a plurality of GPU BOXs connected to a plurality of first communication ports on the dual-node server panel includes: The identity information generated and sent by each second communication port in a plurality of GPU BOXs connected to a plurality of first communication ports on the dual-node server panel after performing resistor pull-up and pull-down processing on a preset number of signal pins of the second communication port is respectively obtained.

10. The multi-GPU BOX management method according to claim 1, characterized in that: The determining the total number of the GPU BOXs and the connection mode between each GPU BOX and the first communication port based on the plurality of identity identification information includes: Sorting the plurality of identity identification information corresponding to the GPU BOX based on the order of the plurality of first communication ports to obtain sorted identification information; The total number of the plurality of GPU BOXs and the connection mode between each of the GPU BOXs and the first communication port are determined based on the sorted identification information.

11. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instruction is executed by a processor, the multi-GPU BOX management method according to any one of claims 1 to 10 is implemented.

12. An electronic device, characterized in that: The method comprises a processor and a memory; wherein the processor implements the multi-GPU BOX management method according to any one of claims 1 to 10 when executing the computer program stored in the memory.

13. A computer-readable storage medium, characterized in that: Used to store computer programs; wherein, when the computer program is executed by a processor, the multi-GPU BOX management method according to any one of claims 1 to 10 is implemented.

Images