Cluster management method, computing server, switch, and cluster system

Abstract

The present invention relates to a cluster management method, a computing server, a switch, and a cluster system. A cluster comprises a cluster management server and a plurality of nodes, wherein each node comprises a plurality of computing servers and at least one switch, and each computing server comprises a plurality of computing devices. The cluster management method comprises: the cluster management server receives respective IP addresses of the computing servers and the switches, and assigns server IDs to the computing servers; each computing server writes its own server ID into the computing devices of the computing server; the switch reads, from the computing devices connected thereto, the server IDs of the computing servers in the same node, and sends the server IDs of the computing servers in the same node to the cluster management server; and the cluster management server establishes a mapping relationship between at least two of the server IDs of the computing servers in the same node, the IP addresses of the computing servers, and the IP address of the switch.

Description

Cluster management method, computing server, switch and cluster system

[0001] Cross-reference to related patent applications

[0002] This application claims priority to and the benefit of Chinese Patent Application No. 202411323521.0, filed September 20, 2024, entitled “Cluster management method, computing server, switch and cluster system,” the entire contents of which are incorporated herein by reference. TECHNICAL FIELD

[0003] The present application relates to the field of communications, and more particularly, to a cluster management method, a computing server, a switch and a cluster system. BACKGROUND

[0004] In a data center for example, computing chips such as GPUs can be used as computing devices, and the interconnection and communication of GPUs are very important, especially the cross-server interconnection and communication of GPUs. SUMMARY

[0005] According to an aspect of the present application, a method for cluster management is provided, the cluster comprising a cluster management server and a plurality of nodes, each of the nodes comprising a plurality of computing servers and at least one switch, each of the computing servers comprising a plurality of computing devices; for at least one of the nodes, the method comprising:

[0006] receiving, by the cluster management server, an IP address of each of the computing servers of the node;

[0007] assigning, by the cluster management server, a server ID to each of the computing servers of the node;

[0008] obtaining a mapping relationship between the server ID of each of the computing servers and the IP address of the computing server.

[0009] In some embodiments, the method further comprises: receiving, by the cluster management server, an IP address of each of the switches in the node; receiving, by the cluster management server, a server ID of a computing server sent by each of the switches in the node; and establishing, by the cluster management server, a mapping relationship between at least two of the server ID of the computing server, the IP address of the computing server and the IP address of the switch within the same node.

[0010] In some embodiments, the method further comprises: performing, by the cluster management server, a topology check according to the server ID of the computing server and the IP address of the corresponding switch within the same node.

[0011] In some embodiments, the method further comprises: the cluster management server sending, to each computing server in the same node, at least one of a server ID and an IP address of each computing server in the node, and an IP address of the switch.

[0012] According to an aspect of the present application, a cluster management server is provided, comprising a memory and a processor, the memory storing computer readable instructions, and the processor executing the computer readable instructions to perform the method.

[0013] According to an aspect of the present application, a method for cluster management is provided, the cluster comprising a cluster management server and a plurality of nodes, each of the nodes comprising a plurality of computing servers and at least one switch, each of the computing servers comprising a plurality of computing devices, the method comprising, for at least one of the nodes:

[0014] each of the computing servers of the node sending an IP address to the cluster management server;

[0015] each of the computing servers of the node receiving a server ID assigned to it by the cluster management server;

[0016] each of the computing servers of the node writing its server ID to each of the computing devices of the computing server.

[0017] In some embodiments, the method further comprises: each of the computing servers of the node receiving at least one of a server ID and an IP address of each computing server in the same node, and an IP address of at least one of the switches, sent by the cluster management server.

[0018] According to an aspect of the present application, a computing server is provided, comprising a memory and a processor, the memory storing computer readable instructions, and the processor executing the computer readable instructions to perform the method.

[0019] According to an aspect of the present application, a method for cluster management is provided, the cluster comprising a cluster management server and a plurality of nodes, each of the nodes comprising a plurality of computing servers and at least one switch, each of the computing servers comprising a plurality of computing devices, the method comprising:

[0020] each of the switches sending an IP address of itself to the cluster management server;

[0021] each of the switches obtaining, from each computing device of the computing server, a server ID of the computing server in which the computing device is located;

[0022] Each of the switches sends the server IDs of the computing servers in the same node to the cluster management server.

[0023] In some embodiments, the obtaining the server IDs of the computing servers comprises traversing each computing device of the node to obtain, from each computing device of the node, a server ID of a computing server on which the computing device is located.

[0024] According to an aspect of the present application, a switch is provided, comprising a memory and a processor, the memory storing computer readable instructions, the processor executing the computer readable instructions to perform the method.

[0025] According to an aspect of the present application, a cluster system is provided, comprising a cluster management server and a plurality of nodes, wherein each of the nodes comprises a plurality of computing servers and at least a switch according to an embodiment of the present application, each of the computing servers comprising a plurality of computing devices.

[0026] According to an aspect of the present application, a cluster system is provided, wherein the computing servers and the switches communicate with each other through optical signals.

[0027] According to an aspect of the present application, a method for cluster management is provided, the cluster comprising a cluster management server and a plurality of nodes, each of the nodes comprising a plurality of computing servers and at least a switch, each of the computing servers comprising a plurality of computing devices, the method comprising, for at least one of the nodes:

[0028] The cluster management server receives IP addresses of the plurality of computing servers and at least one switch, and assigns server IDs to the plurality of computing servers;

[0029] The plurality of computing servers write their server IDs to each computing device of the computing server;

[0030] At least one of the switches reads server IDs of the computing servers in the same node from each computing device, and sends the server IDs of the computing servers in the same node to the cluster management server;

[0031] The cluster management server establishes a mapping relationship between at least two of the server IDs of the computing servers in the same node, the IP addresses of the computing servers, and the IP addresses of the switches.

[0032] In some embodiments, the method further comprises: the cluster management server performing a topology check according to the server IDs of the computing servers in the same node and the IP addresses of the corresponding switches.

[0033] In some embodiments, the method further comprises: sending, by the cluster management server, the server ID and IP address of each computing server in the same node and the IP address of the switch to each computing server in the node.

[0034] In some embodiments, the computing device comprises at least one of a graphics processing unit (GPU), a neural processing unit (NPU), a tensor processing unit (TPU), an intelligent processing unit (IPU), and a deep learning processing unit (DPU).

[0035] According to the embodiments of the present application, the cluster management server aggregates the server ID and IP address of the computing servers in the node and the IP address of the switch, so that it can be determined to which node the computing server belongs, and the relevant information is used for subsequent communication of the computing server and the computing device therein. Specifically, through the cluster management server, the computing server and the switch can obtain the IP address of the other party, and the computing server can also obtain the information of the servers in the same node, including the server ID, IP address, etc.

[0036] It can be seen that the technical solution of the present application solves the problem that the cluster management server, the computing server and the switch cannot automatically identify the roles and exchange information in the initialization stage of the communication system such as the southbound cluster, prepares for the topology collection of the subsequent communication, plays the advantages of the southbound cluster machine hardware, and greatly improves the performance of data transmission.

[0037] The various aspects, features, advantages of the embodiments of the present application will be described in detail below with reference to the accompanying drawings. The above aspects, features, advantages of the present application will become more apparent according to the following specific description in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0038] Fig. 1 is a block diagram showing an exemplary composition of the cluster system of the embodiments of the present application.

[0039] Fig. 2 is a block diagram showing an exemplary composition of the node of the embodiments of the present application.

[0040] Fig. 3 is a schematic diagram showing an exemplary process of the cluster system of the embodiments of the present application. DETAILED DESCRIPTION

[0041] Example embodiments of the present application will be described below in greater detail with reference to the accompanying drawings. The present application may, however, be embodied in many different forms and should not be construed as being limited to the embodiments set forth herein. Rather, these embodiments are provided as exemplary so that the disclosure will be thorough and complete, and fully convey the scope of the various aspects and features of the present application to those skilled in the art. It should be understood that processes, methods, and elements not necessary to the understanding of the present application by persons of ordinary skill in the art can not be described. It should also be understood that features of some embodiments are generally applicable to other embodiments.

[0042] In a cluster such as a data center, computing chips such as GPUs can be used as computing devices, and the interconnection of the GPUs is very important, especially the interconnection of the GPUs in different servers.

[0043] The inventors have found that a southbound interconnection cluster solution can be used, and the communication between the computing devices (e.g., GPUs) in the same node is implemented through a switch, and the performance can be greatly improved.

[0044] However, if the southbound interconnection cluster solution is used, the initialization phase needs to be solved, and the cluster management server, the computing server, and the switch cannot automatically identify the roles and exchange the respective role information.

[0045] FIG. 1 shows an exemplary architecture of a cluster system according to the present application. The cluster system (which can also be referred to simply as a “cluster”) includes a cluster management server 200 and a plurality of nodes 100. Although four nodes are shown in FIG. 1, the present application is not limited thereto. The cluster system of the present application can include any number of nodes as needed. In some embodiments, the cluster management server 200 is a standalone server for managing the topology of the entire cluster. As for the nodes 100, they are nodes in which the computing devices can communicate without network protocol communication (e.g., TCP / IP network protocol, etc.). As shown in FIG. 2, each node 100 includes a plurality of computing servers 300 and at least one switch 400, each computing server 300 includes a plurality of computing devices, and eight GPUs are shown as eight computing devices in FIG. 2, which can be referred to as a group of computing devices. In some embodiments, the cluster management server 200, the computing server 300, and the switch 400 can communicate with each other through a network.

[0046] In the process of performing computation by the cluster, the communication between the computing servers 300 and the switch 400 in the node 100 shown in FIG. 2 includes transmission of optical signals, which are optical signals converted from PCIe electrical signals. Thus, the computing devices in different computing servers in the node 100 can communicate (transmit computing information) without using network protocols, for example, the communication between a computing device 302 in the computing server 300A and a computing device 302 in the computing server 300B can be performed without using network protocols.

[0047] In the process of performing computation by the cluster, in some embodiments, the switch 400 has a switch chip 401 that performs PCIe switching, which switches the path of PCIe electrical signals, and in addition, the switch chip 401 inputs and outputs electrical signals; the switch 400 can further include one or more optical modules to implement the functions of electrical-optical conversion and optical-electrical conversion. In some embodiments, the switch 400 can be an all-optical switch, which can switch the optical path by the switch chip 401, i.e., the switch chip 401 inputs and outputs optical signals, which switches the path of optical signals to achieve switching, so that the transmission of computing information between the computing devices can be performed in the process of performing computation by the cluster.

[0048] Although FIG. 2 shows four computing servers (including the computing server 300A, the computing server 300B, the computing server 300C, and the computing server 300D), each of which has eight computing devices 302, e.g., eight GPUs, it should be understood that the present application is not limited thereto. The computing servers can be referred to as GPU servers because they include eight GPUs (eight computing chips), although the GPU servers can include an appropriate number of CPUs (not shown in FIG. 2) as well, in which case, a CPU+GPU heterogeneous computing architecture (heterogeneous computing) can be implemented within a GPU server. In FIG. 2, each computing server and the GPUs therein are connected to the switch 400, and the computing devices in different computing servers can be connected through the switch 400. More or fewer computing servers can be provided in a node 100 as needed. In addition, two or more switches 400 can be configured in a node 100 as needed according to the communication bandwidth and other communication connection requirements. For example, the node 100 can include two switches, e.g., the switch 400 (shown in FIG. 2) as a first switch, and a second switch, each computing server and each computing device being connected to the second switch, and the computing devices in different computing servers being connected through the second switch.

[0049] Exemplarily, the computing device can include a computing chip, for example, the computing chip can be a GPU (for example, shown in FIG. 2), and one computing chip (for example, one GPU) can include a plurality of computing cores.

[0050] The computing device can include other types of computing chips in addition to the GPU. The computing chip can include but is not limited to a graphics processing unit (GPU), a neural network processing unit (NPU), a tensor processing unit (TPU), an intelligent processing unit (IPU), a deep learning processing unit (DPU), etc., which can be used for artificial intelligence computing, referred to as an artificial intelligence chip.

[0051] The computing device can be a computing card (computing board card) with a computing chip, which can be used to accelerate computing, also known as an acceleration card. For example, for a GPU, it can be packaged and manufactured as a computing board card containing the GPU, and the eight GPUs in the same computing server in FIG. 2 can be packaged and manufactured as eight corresponding computing board cards, respectively, and the eight computing board cards can be plugged into the circuit substrate of the computing server.

[0052] In an exemplary embodiment, a client / server architecture is adopted, the cluster management server 200 is the server, and the computing servers 300A, 300B, 300C, 300D and the switch 400 are the clients. In general, the client obtains the ID assigned by the server from the server, and transmits the IP address and other information to the server, and the server acts as a medium to distribute the information. For the present application, the computing servers 300A, 300B, 300C, 300D write the information such as the server ID of each server and the ID of each GPU to the storage space of the GPU connected to the switch 400. The switch 400 reads the written information from the storage space and transmits it to the cluster management server 200. The cluster management server 400 aggregates the information, and the subsequent related information can be used to determine to which node the computing server 300 belongs.

[0053] In an exemplary embodiment, the node management module of each node is deployed on each computing server, specifically, each computing server can include a node management module 301, which is responsible for collecting topology information and using the collected information for subsequent communication. In some embodiments, the node management module 301 includes a node management program.

[0054] According to FIG. 2, the computing device is a GPU, so the ID of the GPU mentioned in the context is also the ID of the computing device, and the port ID of the GPU is also the port ID of the computing device.

[0055] In an exemplary embodiment, as shown in FIG. 3, the cluster management method for the cluster system comprises: S1, the computing server 300 acquires the server ID assigned by the cluster management server 200; S2, the computing server 300 sends its IP address to the cluster management server 200. Wherein, the computing server 300 comprises a node management module 301 and a computing device 302 (e.g. GPU), and the cluster management server 200 assigns a server ID for each computing server 300. Exemplarily, the computing server 300 can communicate with the cluster management server through the IP address of the cluster management server, or the host name (HOST NAME) of the cluster management server. S3, the switch 400 also sends its IP address to the cluster management server 200. Thus, the mapping relationship between the server ID and the IP address of the computing server can be obtained. Exemplarily, the cluster management server 200 can establish the mapping relationship between the server ID and the IP address of the computing server, for example, obtain a two-dimensional table of the server ID and the IP address of the computing server. In some embodiments, the cluster management server 200 is configured to dynamically assign the server ID of the computing server, and the information of the computing server can also be dynamically determined, so that the computing server can be replaced flexibly, for example, after replacing a new computing server, the related server ID information can also be dynamically updated, etc.

[0056] Back to FIG. 3, the cluster management method further comprises: S4, the computing server 300 writes its server ID, the GPU ID of each GPU contained therein and the port ID of the corresponding GPU into the GPU (e.g. the storage space thereof). In some embodiments, the computing server 300 writes the server ID, the GPU ID and the port ID of the GPU into each GPU (computing device 302) through the node management module 301, for example, writes into the storage space of each GPU (computing device). The node management module 301 uses the interface provided by the GPU to write the server ID, the GPU ID and the port ID of the GPU into the reserved storage space. In some embodiments, the node management module 310 comprises a node management program, which is on the same computing server 300 as the GPU 302, and the writing operation is realized through software function call, etc.

[0057] Back to FIG. 3, the cluster management method further comprises: S5, the switch 400 acquires the server ID, the GPU ID and the port ID of the GPU from the computing server 300. In some embodiments, the switch 400 reads the corresponding storage space of each GPU 302 connected thereto, so as to obtain the server ID, the GPU ID and the port ID of all the computing servers 300 in the same node.

[0058] Then, the cluster management method further comprises: S6, the switch 400 sends the server ID of the computing server in the same node to the cluster management server 200. In addition, the cluster management server gets the IP address of all computing servers and the IP address of the switch, thereby being able to deduce which computing servers belong to the same node, and can establish the mapping relationship between at least two of the server ID of the computing server, the IP address of the computing server and the IP address of the switch in the same node. For example, the mapping relationship between the server ID of the computing server and the IP address of the switch; for example, the mapping relationship between the IP address of the computing server and the IP address of the switch; for example, the mapping relationship between the server ID of the computing server and the IP address of the computing server; or the above-mentioned relationship among the server ID of the computing server, the IP address of the computing server and the IP address of the switch.

[0059] By traversing each node, the above operation can be performed on each node in the cluster to obtain the mapping relationship between at least two of the server ID of the computing server, the IP address of the computing server and the IP address of the switch in each node.

[0060] Further, the cluster management method comprises: S7, the cluster management server 200 performs topology checking according to the server ID of the computing server in the same node and the IP address of the corresponding switch, for example, checks the connection of the computing server in the node, the connection of the switch, whether they are correctly connected, and the connection relationship between the computing server and the switch. In this way, once a new computing server is connected or connected incorrectly in the node, the cluster management server 200 can discover it in time according to the mapping relationship and take processing measures.

[0061] In addition, the cluster management method further comprises: S8, the cluster management server 200 sends the server ID and IP address of the computing server in the same node, the IP address of the switch to each computing server 300 in the node. For example, the cluster management server 200 sends the summarized information (including the server ID and IP address of the computing server in the same node, the IP address of the switch, etc.) to each computing server 300, for example, to the node management module 301 of the server 300, and the computing server 300 thereby gets the IP address of the corresponding switch 400, and can further communicate with the corresponding switch based on the IP address. In some embodiments, the node management module 301 communicates with the switch 400 to obtain topology information, and then uses the topology information for subsequent communication, for subsequent communication of the cluster.

[0062] It can be learned from the above description that the embodiment of the application solves the problem that the cluster management server, the computing server and the switch cannot automatically identify roles and exchange information in the southbound cluster initialization stage, is beneficial to subsequent topology collection of the cluster, plays the advantages of hardware of the southbound cluster machine and greatly improves the performance of data transmission.

[0063] In some embodiments, the method further comprises: the cluster management server sending the server ID and the IP address of each computing server in the same node and the IP address of the switch to each computing server in the node.

[0064] In some embodiments, the method further comprises: the cluster management server performing topology checking according to the server ID of the computing server in the same node and the IP address of the corresponding switch.

[0065] In some embodiments, the method further comprises: the cluster management server sending the server ID and the IP address of each computing server in the same node and the IP address of the switch to each computing server in the node.

[0066] In some embodiments, the switch reads the server ID of the computing server in the same node from the GPU (computing device) connected thereto.

[0067] In some embodiments, the method further comprises: the cluster management server sending the server ID and the IP address of each computing server in the same node and the IP address of the switch to each computing server in the node.

[0068] In some embodiments, the method further comprises: the cluster management server receiving an IP address of each switch; the cluster management server receiving a server ID of a computing server sent by each switch in the node; and the cluster management server establishing a mapping relationship between at least two of the server ID of the computing server, the IP address of the computing server, and the IP address of the switch in the same node.

[0069] In some embodiments, the method further comprises: the cluster management server performing a topology check according to the server ID of the computing server and the IP address of the corresponding switch in the same node.

[0070] In some embodiments, the method further comprises: the cluster management server sending the server ID and the IP address of each computing server, and the IP address of the switch in the same node to each computing server in the node.

[0071] Accordingly, embodiments of the present application also relate to a cluster management server comprising a memory and a processor, the memory storing computer readable instructions, and the processor executing the computer readable instructions to perform the method of any of the above embodiments.

[0072] In some other embodiments, the present application provides a method for cluster management, the cluster comprising a cluster management server and a plurality of nodes, each of the nodes comprising a plurality of computing servers and at least one switch, each of the computing servers comprising a plurality of computing devices. For at least one of the nodes, the method comprises: each of the computing servers in the node sending an IP address to the cluster management server; each of the computing servers in the node receiving a server ID assigned by the cluster management server; and each of the computing servers in the node writing its own server ID to each of the computing devices in the computing server. In some embodiments, the computing devices comprise GPUs, and the computing server writes its own server ID, GPU ID, and port ID of the GPU to a storage space of the corresponding GPU port.

[0073] In some embodiments, the method further comprises: each of the computing servers in the node receiving the server ID and the IP address of the computing servers in the same node, and the IP address of the switch sent by the cluster management server.

[0074] Accordingly, embodiments of the present application also relate to a computing server comprising a memory and a processor, the memory storing computer readable instructions, and the processor executing the computer readable instructions to perform the method of any of the above embodiments.

[0075] In some embodiments, the method further comprises: the switch sending an IP address of the switch itself to the cluster management server; the switch obtaining server IDs of the computing servers in the same node from the computing devices of the computing servers; and the switch sending the server IDs of the computing servers in the same node to the cluster management server.

[0076] In some embodiments, the switch obtaining the server IDs of the computing servers in the same node comprises: traversing each computing device of the node to obtain, from each computing device, a server ID of a computing server to which the computing device belongs. Thus, the switch obtains the server IDs of each computing server in the same node.

[0077] In some embodiments, the switch reads the server IDs of the computing servers in the same node, GPU IDs and port IDs of the GPUs from a storage space of at least one GPU (computing device) of the computing servers. Further, the switch traverses each GPU (computing device) of the node to obtain the server IDs of the computing servers in the same node, the GPU IDs (computing device IDs) and the port IDs of the GPUs (computing device port IDs).

[0078] Accordingly, embodiments of the present application also relate to a switch comprising a memory and a processor, the memory storing computer readable instructions, and the processor executing the computer readable instructions to perform the above method.

[0079] Those skilled in the art will understand that, although the present application and its embodiments are proposed to solve the problem that the cluster management server, the computing servers and the switch cannot automatically identify roles and exchange information in the initialization stage of the southbound interconnection cluster, the present application can be applied to other communication systems that need to automatically identify roles and exchange information.

[0080] Those skilled in the art will understand that the above embodiments are only for illustrative purposes and cannot limit the scope of the patent protection claimed by the present application. Modifications, changes, equivalent replacements, etc. made in accordance with the above embodiments still fall within the scope covered by the claims of the present application.

Claims

1. A method for cluster management, the cluster comprising a cluster management server and a plurality of nodes, each of the nodes comprising a plurality of computing servers and at least one switch, each of the computing servers comprising a plurality of computing devices; For The method comprises, for at least one of the nodes: The cluster management server receives the IP address of each computing server in the node; The cluster management server assigns a server ID to each computing server in the node; The cluster management server obtains a mapping relationship between the server ID of each computing server and the IP address of the computing server.

2. The method of claim 1, wherein, Further comprising: The cluster management server receives the IP address of each switch in the node; The cluster management server receives the server ID of a computing server sent by each switch in the node; The cluster management server establishes a mapping relationship between at least two of the server ID of a computing server, the IP address of the computing server, and the IP address of the switch in the same node.

3. The method of claim 2, wherein, Further comprising: The cluster management server performs a topology check according to the server ID of a computing server and the IP address of the corresponding switch in the same node.

4. The method of claim 2, wherein, Further comprising: The cluster management server sends the server ID and IP address of each computing server, and the IP address of the switch in the same node to each computing server in the node.

5. A cluster management server comprising a memory and a processor, the memory storing computer-readable instructions, and the processor executing the computer-readable instructions to perform the method of any one of claims 1 to 4.

6. A method for cluster management, the cluster comprising a cluster management server and a plurality of nodes, each of the nodes comprising a plurality of computing servers and at least one switch, each of the computing servers comprising a plurality of computing devices, the method comprising, for at least one of the nodes: Each of the computing servers in the node sends an IP address to the cluster management server; Each of the computing servers in the node receives a server ID assigned by the cluster management server; Each of the computing servers in the node writes its own server ID to each of the computing devices in the computing server.

7. The method of claim 6, wherein, Further comprising: Each of the computing servers in the node receives at least one of the server ID and IP address of each computing server in the same node, and the IP address of at least one of the switches, sent by the cluster management server.

8. A computing server comprising a memory and a processor, the memory storing computer-readable instructions, and the processor executing the computer-readable instructions to perform the method of claim 6 or 7.

9. A method for cluster management, the cluster comprising a cluster management server and a plurality of nodes, each of the nodes comprising a plurality of computing servers and at least one switch, each of the computing servers comprising a plurality of computing devices, the method comprising: Each of the switches sends its own IP address to the cluster management server; Each of the switches obtains the server ID of the computing server in which the computing device is located from each computing device of the computing server; Each of the switches sends the server ID of the computing server in the same node to the cluster management server.

10. The method of claim 9, wherein, The obtaining the server ID of the computing server in which the computing device is located comprises: Traverse each computing device of the node to obtain a server ID of a computing server where each computing device is located from each computing device of the node.

11. A switch comprising a memory storing computer-readable instructions and a processor executing the computer-readable instructions to perform the method of claim 9 or 10.

12. A cluster system comprising the cluster management server of claim 5 and a plurality of nodes, wherein, Each of the nodes comprises a plurality of computing servers as claimed in claim 8 and at least one switch as claimed in claim 11, each computing server comprising a plurality of computing devices.

13. The cluster system of claim 12, wherein, The computing servers and the switches communicate with each other through optical signals.

14. A method for cluster management, the cluster comprising a cluster management server and a plurality of nodes, each of the nodes comprising a plurality of computing servers and at least one switch, each computing server comprising a plurality of computing devices, the method comprising, for at least one of the nodes: The cluster management server receives IP addresses of the plurality of computing servers and at least one switch, and assigns server IDs to the plurality of computing servers; The plurality of computing servers write their server IDs to each computing device of the computing server; At least one of the switches reads server IDs of computing servers in the same node from each computing device, and sends the server IDs of the computing servers in the same node to the cluster management server; The cluster management server establishes a mapping relationship between at least two of the server IDs of the computing servers in the same node, the IP addresses of the computing servers, and the IP address of the switch.

15. The method of claim 14, wherein, Further comprising: The cluster management server performs a topology check according to the server IDs of the computing servers in the same node and the IP address of the corresponding switch.

16. The method of claim 14, wherein, Further comprising: The cluster management server sends the server IDs and the IP addresses of the computing servers in the same node, and the IP address of the switch to each computing server of the node.

17. The method of any of claims 1-6, 9-10, 14-16, the computing device comprising at least one of a graphics processing unit (GPU), a neural processing unit (NPU), a tensor processing unit (TPU), an intelligent processing unit (IPU), and a deep learning processing unit (DPU).

Images