Server cabinet and server system

Abstract

The invention discloses a server cabinet and a server system, and relates to the technical field of servers, the server cabinet comprises a computing connector, a first switching connector and a second switching connector, the computing connector is divided into a first pin set and a second pin set; a first connection relation exists between the first pin set and the first switching connector, and the first connection relation is matched with processor information of the first processor with the target bandwidth; a second connection relation exists between the second pin set and the second switching connector, and the second connection relation is matched with the bandwidth difference value and the first connection relation. Through the method and the device, the technical problem that the utilization rate of the connection resources of the server cabinet in the related technology is relatively low can be solved, and the technical effect of improving the utilization rate of the connection resources of the server cabinet is achieved.

Description

Technical Field

[0001] The present application relates to the field of server technology, and in particular to a server cabinet and a server system. Background Art

[0002] The server cabinet is usually used to provide a communication environment for the server processor deployed in the server through the exchange processor, so that the server processor can realize the server function. The current server cabinet is usually designed according to the specifications of the server processor supported by the server. As the specifications of the server processor supported by the server cabinet are no longer limited to one, the use of a fixed connection form of the server cabinet to cope with server processors of various specifications is likely to cause a waste of the connection resources of the server cabinet. Summary of the invention

[0003] The present application provides a server cabinet and a server system to at least solve the problem of low utilization rate of connection resources of the server cabinet in the related art.

[0004] The present application provides a server cabinet, which includes: a computing connector, a first switching connector, and a second switching connector, the computing connector is divided into a first pin set and a second pin set; a first connection relationship is provided between the first pin set and the first switching connector, the first connection relationship matches the processor information of a first processor having a target bandwidth, wherein the first pin set corresponds to a signal of the target bandwidth; a second connection relationship is provided between the second pin set and the second switching connector, the second connection relationship matches the bandwidth difference and the first connection relationship, wherein the bandwidth difference is the difference between a reference bandwidth of the second processor and the target bandwidth, and the second pin set corresponds to a signal of the bandwidth difference.

[0005] The present application also provides a server system, which includes: a computing node, a switching node and the aforementioned server cabinet, the computing connector included in the server cabinet is connected to the server processor in the computing node, and the switching connector included in the server cabinet is connected to the switching processor in the switching node.

[0006] Through the present application, the computing connector in the server cabinet is divided into a first pin set corresponding to the signal of the target bandwidth and a second pin set corresponding to the signal of the bandwidth difference. The first connection relationship between the first pin set and the first switching connector matches the processor information of the first processor with the target bandwidth, and the second connection relationship between the second pin set and the second switching connector matches the bandwidth difference between the reference bandwidth and the target bandwidth of the second connector and the aforementioned first connection relationship. When dealing with server processors with different bandwidths, the server cabinet can provide a connection method that matches the connected server processor, avoiding the waste of connection resources of the server cabinet caused by using a fixed connection method. Therefore, the technical problem of low utilization of connection resources of the server cabinet in the related technology can be solved, and the technical effect of improving the utilization of connection resources of the server cabinet can be achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] In order to more clearly illustrate the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0008] Figure 1 is a schematic diagram of the front of a cabinet of an AI super-node whole-cabinet server according to an embodiment of the present application;

[0009] Figure 2 is a schematic diagram of the back of a cabinet of an AI super-node whole-cabinet server according to an embodiment of the present application;

[0010] Figure 3 Schematic diagram of the connection between the computing node and the switching node of an AI super node whole cabinet server according to an embodiment of the present application Figure 1 ;

[0011] Figure 4 Schematic diagram of the connection between the computing node and the switching node of an AI super node whole cabinet server according to an embodiment of the present application Figure 2 ;

[0012] Figure 5 Schematic diagram of the connection between the computing node and the switching node of an AI super node whole cabinet server according to an embodiment of the present application Figure 3 ;

[0013] Figure 6 is a schematic diagram of a server cabinet according to an embodiment of the present application;

[0014] Figure 7 is a connection diagram of a computing connector according to an embodiment of the present application;

[0015] Figure 8 is a connector layout diagram according to an embodiment of the present application;

[0016] Fig. 9 is a schematic diagram of a server system according to an embodiment of the present application;

[0017] Fig.10 is a schematic diagram of a ratio between the bandwidth of a large-bandwidth GPU and the bandwidth of a small-bandwidth GPU according to an embodiment of the present application;

[0018] Fig.11 A schematic diagram of a connection relationship between a computing node and a switching node according to an embodiment of the present application Figure 1 ;

[0019] Fig.12 A schematic diagram of a connection relationship between a computing node and a switching node according to an embodiment of the present application Figure 2 . DETAILED DESCRIPTION

[0020] 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.

[0021] It should be noted that, in the description of this application, the terms "include", "comprise" or any other variant 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 includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. The terms "first", "second", etc. in this application are used to distinguish similar objects, and are not used to describe a specific order or sequence.

[0022] With the continuous development of artificial intelligence technology, AI (Artificial Intelligence) model training and reasoning have attracted more and more attention. As the key infrastructure supporting AI model training and reasoning, the new generation of computing cluster architecture is facing severe performance challenges. The marginal benefits of single server processor performance are decreasing, approaching the physical limit, and it is difficult to meet the continuous growth demand for computing power of large AI models. At the same time, the horizontal expansion strategy of computing clusters by simply increasing the number of servers has also encountered bottlenecks in efficiency and scalability, and faces challenges including cost, data synchronization and energy consumption, which hinders the efficient execution of large-scale parallel computing. In this context, the AI ​​super-node whole cabinet system came into being. It systematically integrates multiple groups of GPU (Graphics Processing Unit) computing units and switching units through an innovative topological architecture, and comprehensively optimizes the integration of multiple computing power, high-speed advanced interconnection, heat dissipation, high power density power supply, and whole cabinet management. It provides a scalable high-bandwidth domain (HBD) super node system with 32 cards or more to meet the cutting-edge large model training and reasoning needs.

[0023] Figure 1 is a schematic diagram of the front of a cabinet of an AI super node whole cabinet server according to an embodiment of the present application. Figure 1 As shown, the number of Us represents the number of layers of the whole cabinet server cabinet. The height of each U position can be adjusted. Generally, the square hole bar of each node of the cabinet can be set to OU (48mm), RU (44.45mm) or SU (46.5mm) in three specifications, but not limited to. The various components of the server can be deployed in each layer of the whole server cabinet in a dispersed manner, for example, the management switch can be deployed at the 44OU and 42OU positions, the cable tray can be deployed at the 43OU position, the Powershelf (power rack, used to provide centralized power management for devices in the server rack) node can be deployed at the 37~38OU position and the 5~6OU position, the Rack Stiffener (rack reinforcement, used to increase the structural strength and stability of the server rack) node can be deployed at the 36OU position and the 7OU position, and the CDU (Cooling Distribution Unit, a cooling distribution unit used to manage and distribute cooling liquid or cold air) node can be deployed at the 1~4OU position. Figure 1 In the distribution shown, the OU positions 39 to 41 have not yet been considered for deployment of any components, so they are marked as NA (Not Applicable) layers. In actual applications, specific functional components can be deployed in the NA layer based on the functional requirements of the server. The computing nodes are placed on the upper and lower sides of the cabinet (for example Figure 1The 28 OU to 35 OU and 8 OU to 15 OU shown in the figure are used to deploy computing nodes. The switching nodes are placed in the middle of the cabinet (for example Figure 1 The 16 OU to 27 OU shown in the figure are deployed with switching nodes). This design can make the deployment of power cables and communication cables more balanced. Figure 1 The configuration shown in FIG. 1 is only a typical configuration, and the number of switching nodes and the number of computing nodes can be adjusted according to actual conditions.

[0024] Figure 2 Schematic diagram of the back of a cabinet of an AI super node whole cabinet server according to an embodiment of the present application. Figure 2 As shown, the back of the whole cabinet server mainly includes three components: Busbar, Manifold and Cable Tray. The Busbar is located in the middle (or one side), and the Powershelf node is connected to the Busbar to power the whole cabinet. The Cable Tray is located on the left and right sides of the Busbar. The Cable Tray is used to connect the high-speed signals of the computing nodes and the switching nodes according to the designed topological interconnection relationship. The Manifold is located on the left and right sides of the rear of the cabinet, providing water inlet and outlet, and providing a liquid path for liquid cooling for the entire cabinet. These components work together to ensure that the power supply and cooling system of the whole cabinet server operate efficiently and safely.

[0025] Figure 3 Schematic diagram of the connection between the computing node and the switching node of an AI super node whole cabinet server according to an embodiment of the present application Figure 1 .like Figure 3As shown, #1~#N are N computing nodes, and one CPU (Central Processing Unit) can be deployed on each computing node, but is not limited to, and the CPU can be but is not limited to hanging 4 GPUs through the PCIe (Peripheral Component Interconnect Express, a high-speed serial computer expansion bus standard) switching chip. Taking 16 computing nodes deployed in the AI ​​whole cabinet server as an example, there are 16×4=64 GPUs in this AI whole cabinet server. If N computing nodes are deployed in the AI ​​whole cabinet server, there are 4×N GPUs in this AI whole cabinet server. The computing node can be but is not limited to also be deployed with a DPU (Data Processing Unit) network card and storage device and a BMC (Baseboard Management Controller) management module. The storage device can be but is not limited to an SSD (Solid State Drive). Figure 3 The switching nodes 1 to M shown are M switching nodes, and each switching node may have, but is not limited to, a MAC (Media Access Control) chip. The high-speed signals of the switching nodes and computing nodes are interconnected through the Cable Tray. In the AI ​​whole-rack server, all computing nodes and switching nodes are connected to the Rack Cable Tray (cable tray in or near the rack) through high-density connectors to achieve high-speed Serdes (Serializer / Deserializer) interconnection between all GPUs in the AI ​​whole-rack server. Figure 4 Schematic diagram of the connection between the computing node and the switching node of an AI super node whole cabinet server according to an embodiment of the present application Figure 2 , the Serdes high-speed interconnection method between all the aforementioned GPUs is as follows Figure 4 It should be noted that the solution of fully interconnecting computing nodes in the AI ​​whole cabinet server through switching nodes requires that the high-speed signal resources of each GPU in each computing node be evenly distributed to each switching node. This is a scale-up topology. Figure 5 Schematic diagram of the connection between the computing node and the switching node of an AI super node whole cabinet server according to an embodiment of the present application Figure 3 ,like Figure 5As shown in the figure, computing nodes 1-8 are interconnected through switching nodes 1-8. The connector of each computing node in the AI ​​whole cabinet server is connected to 4 GPUs, and the signals at the same position of 8 computing nodes are connected to a switch (i.e., switching node). Figure 5 The computing node signals and switching node signals with the same number shown in FIG. 1 have a connection relationship.

[0026] For the above-mentioned AI whole cabinet server, when there are multiple specifications (i.e., bandwidth) of GPUs that can be connected in the computing nodes, if a fixed connection method of computing nodes and switching nodes is used regardless of the specifications of the GPUs connected to the computing nodes, it will lead to a waste of connection resources of the server cabinet. In order to solve the above problems, the present application proposes a server cabinet and a server system.

[0027] In order to enable those skilled in the art to better understand the present application, the present application is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0028] The embodiment of the present application provides a server cabinet, Figure 6 is a schematic diagram of a server cabinet according to an embodiment of the present application, such as Figure 6 As shown, the server cabinet includes:

[0029] a computing connector, a first switching connector, and a second switching connector, the computing connector being divided into a first pin set and a second pin set;

[0030] A first connection relationship exists between the first pin set and the first switching connector, the first connection relationship matches the processor information of the first processor having the target bandwidth, wherein the first pin set corresponds to a signal of the target bandwidth;

[0031] There is a second connection relationship between the second pin set and the second switching connector, and the second connection relationship matches the bandwidth difference and the first connection relationship, wherein the bandwidth difference is the difference between the reference bandwidth and the target bandwidth of the second processor, and the second pin set corresponds to the signal of the bandwidth difference.

[0032] Through the above content, since the computing connector in the server cabinet is divided into a first pin set corresponding to the signal of the target bandwidth and a second pin set corresponding to the signal of the bandwidth difference, the first connection relationship between the first pin set and the first switching connector matches the processor information of the first processor with the target bandwidth, and the second connection relationship between the second pin set and the second switching connector matches the bandwidth difference between the reference bandwidth and the target bandwidth of the second connector and the aforementioned first connection relationship. Then, when dealing with server processors with different bandwidths, the server cabinet can provide a connection method that matches the connected server processor, avoiding the waste of connection resources of the server cabinet caused by using a fixed connection method. Therefore, the technical problem of low utilization of connection resources of the server cabinet in the related technology can be solved, and the technical effect of improving the utilization of connection resources of the server cabinet can be achieved.

[0033] Optionally, in this embodiment, the computing connector may be, but is not limited to, a connector for connecting a server processor. The server processor may be, but is not limited to, a device with computing functions, and the server processor may be, but is not limited to, a GPU or a CPU, etc. The first switching connector and the second switching connector may be, but are not limited to, connectors for connecting a switching processor, and the switching processor may be, but is not limited to, a device with signal / data switching functions, and the switching processor may be, but is not limited to, a switching chip. The connector may be, but is not limited to, a device that can provide a path for data / signal transmission, and the connector may be, but is not limited to, a PCIe connector or a high-density connector, etc.

[0034] Optionally, in this embodiment, the bandwidth specification of the computing connector can be, but is not limited to, determined based on the maximum processor bandwidth of the server processor that the server cabinet needs to support. For example, when the bandwidth specifications of the server processor that the server cabinet needs to support include bandwidth A, bandwidth B, and bandwidth C, where A>B>C, a connector with a bandwidth of A (or a bandwidth greater than A) needs to be selected as the computing connector of the server cabinet. For another example, when the bandwidth specifications of the server processor supported by the server cabinet include bandwidth D and bandwidth E, where D>E, a connector with a bandwidth of D (or a bandwidth greater than D) needs to be selected as the computing connector of the server cabinet.

[0035] Optionally, in this embodiment, the second switching connector and the first switching connector may have, but are not limited to, the same bandwidth specification.

[0036] Optionally, in this embodiment, the first pin set may be, but is not limited to, used to connect a first processor having a target bandwidth, and a combination of the first pin set and the second pin set may be, but is not limited to, used to connect a second processor having a reference bandwidth. Figure 7is a connection diagram of a computing connector according to an embodiment of the present application. Figure 7 As shown, the connector of the computing node (i.e., the computing connector) is divided into a common bandwidth portion (i.e., a first pin set) corresponding to a small bandwidth (i.e., a target bandwidth) signal and an additional bandwidth portion (i.e., a second pin set) corresponding to an additional bandwidth signal of a large bandwidth (i.e., a reference bandwidth) compared to the small bandwidth. It is possible, but not limited to, to determine the bandwidth of multiple server processors to be connected to the server cabinet by determining the bandwidth of the server processor with the smallest bandwidth as the small bandwidth. In this case, the additional bandwidth is the difference between the bandwidth of the server processors other than the server processor with the smallest bandwidth among the multiple server processors and the small bandwidth. The current additional bandwidth can be determined according to the bandwidth of the server processor currently connected to the server cabinet.

[0037] Optionally, in this embodiment, the first connection relationship, i.e., the connection relationship between the first pin set and the first exchange connector, may include, but is not limited to, a first number of calculation connectors, a second number of first exchange connectors, and a first connection method between the first number of first pin sets and the second number of first exchange connectors, and may be, but is not limited to, characterizing the aforementioned first connection method with a first ratio. The second connection relationship, i.e., the connection relationship between the second pin set and the second exchange connector, may include, but is not limited to, a third number of second exchange connectors and a second connection method between the first number of second pin sets and the third number of second exchange connectors, and may be, but is not limited to, characterizing the aforementioned second connection method with a second ratio.

[0038] Optionally, in this embodiment, the first connection relationship / the second connection relationship may be realized by, but is not limited to, connecting with a Cable Tray (a structure for supporting and managing cables).

[0039] Optionally, in this embodiment, the processor information may be, but is not limited to, used to indicate the signal transmission requirement of the first processor. The processor information may include, but is not limited to, the number of first processors and the bandwidth of the first processor.

[0040] Optionally, in this embodiment, the first connection relationship matches the processor information of the first processor with the target bandwidth, including but not limited to that the first connection relationship is established according to the processor information. The first connection relationship matches the processor information, but is not limited to indicating that when the first pin set and the first switching connector are connected according to the first connection relationship, the waste of connector connection resources can be minimized while ensuring that the first processor is fully interconnected.

[0041] Optionally, in this embodiment, the bandwidth difference is the difference between the reference bandwidth of the second processor and the target bandwidth of the first processor. The full - interconnection implementation solution of the server processors within the specified range of the server processors to be connected to the server cabinet can be designed by, but not limited to, regarding server processors with different bandwidths as the second processor. For example, the bandwidth of server processor one, which is the server processor with the minimum bandwidth that the server cabinet can support (or to be connected to), is a, and the bandwidth of server processor two, which is the server processor with the maximum bandwidth that the server cabinet can support (or to be connected to), is b (a < b). Server processor one can be regarded as the first processor, and server processors with bandwidths in the range of a to b (excluding a and including b) can be regarded as the second processor.

[0042] Optionally, in this embodiment, the second connection relationship matching the bandwidth difference and the first connection relationship includes, but is not limited to, the second connection relationship being established based on the bandwidth difference and the first connection relationship. The second connection relationship matching the bandwidth difference and the first connection relationship can, but is not limited to, mean that when connecting the first pin set and the first switching connector according to the first connection relationship and connecting the second pin set and the second switching connector according to the second connection relationship, the waste of connector connection resources can be minimized while ensuring the full - interconnection of the second processor.

[0043] Optionally, in this embodiment, the second ratio included in the second connection relationship between the second pin set and the second switching connector is not limited to only being the first ratio included in the first connection relationship, but can be specifically set according to the bandwidth difference and the first connection relationship, making the topological structure within the server cabinet more flexible.

[0044] Optionally, in this embodiment, the server cabinet can, but is not limited to, be used to establish a full - interconnection architecture for multiple computing nodes included in the server system. For a full - interconnection architecture, a switching node must have a connection relationship with all computing nodes. The server system can, but is not limited to, include computing nodes and switching nodes. Server processors are provided within the computing nodes, and switching processors are provided within the switching nodes. Connectors are provided in the server cabinet corresponding to the switching nodes and the computing nodes. Through the connection of the connectors of the computing nodes and the connectors of the switching nodes, the signals of each computing node are relayed through the switching node to achieve full - interconnection.

[0045] Optionally, in this embodiment, multiple computing connectors are deployed on multiple computing node chassis according to, but not limited to, heat dissipation requirements and power supply requirements, etc. The number of computing connectors deployed on each computing node chassis can, but is not limited to, be the same.

[0046] Optionally, in this embodiment, including but not limited to deploying multiple first switching connectors on multiple switching node chassis according to heat dissipation requirements and power supply requirements, and deploying multiple second switching connectors on other multiple switching node chassis, the number of switching connectors deployed on each switching node chassis can be but is not limited to being the same. In order to simplify the management of the switching node chassis, generally, both the first switching connector and the second switching connector are not deployed on one switching node chassis.

[0047] Optionally, in this embodiment, the switch nodes deployed on each switch node chassis may, but are not limited to, be able to communicate with each other.

[0048] As an optional implementation, the first connection relationship includes: calculating a first quantity of connectors, a second quantity of first switching connectors, and a first ratio, the first ratio being used to indicate a proportional relationship between the first pin set connecting to the first switching connector and a proportional relationship between the first switching connector connecting to the first pin set; the second quantity is calculated based on the first quantity, the target bandwidth, the switch bandwidth of the switching processor to which the first switching connector is to be connected, and the first ratio.

[0049] Optionally, in this embodiment, the first number of computing connectors, that is, the number of all computing connectors in the server cabinet, may be determined, but not limited to, according to the number of first processors (or second processors).

[0050] Optionally, in this embodiment, the second number of first switching connectors is the number of all first switching connectors in the server cabinet.

[0051] Optionally, in this embodiment, the first ratio is used to indicate the proportional relationship between the first pin set and the first switching connector and the proportional relationship between the first switching connector and the first pin set. For example, when the first ratio is 3:3, it means that one first switching connector has a connection relationship with three first pin sets, and one first pin set has a connection relationship with three first switching connectors.

[0052] As an optional embodiment, the server cabinet also includes a first switching node chassis, and the first switching connector is deployed on the first switching node chassis; the number of first switching node chassis is the ratio of the second number to the first unit number, and the first unit number is the number of first switching connectors that each first switching node chassis is allowed to accommodate.

[0053] Optionally, in this embodiment, the first switching connectors may be, but are not limited to, evenly distributed on each first switching node chassis.

[0054] Optionally, in this embodiment, the first unit quantity may be determined from a perspective including but not limited to heat dissipation requirements.

[0055] Through the above content, by clarifying that the number of the first switching node chassis is the ratio of the second number (the number of the first switching connectors) to the first unit number (the number of connectors that each switching node chassis can accommodate), the rationalization of the number of switching node chassis is ensured. This design avoids the waste of switching node chassis resources, because the configuration of each switching node chassis is determined based on the number of connectors that it can effectively carry, thereby improving the overall resource utilization efficiency.

[0056] As an optional implementation, the number of computing connectors is a first number, wherein the first number is determined based on the number of first processors, the server cabinet further includes a computing node chassis, and the computing connectors are deployed on the computing node chassis; the number of computing node chassis is a ratio of the first number to a second unit number, the second unit number being the number of computing connectors that each computing node chassis is allowed to accommodate; each first switching node chassis is connected to all the computing node chassis under a first connection relationship.

[0057] Optionally, in this embodiment, the number of computing connectors may be, but is not limited to, determined according to the number of first processors. For example, the number of computing connectors may be the number of first processors.

[0058] Optionally, in this embodiment, the computing connectors may be, but are not limited to, evenly distributed on each computing node chassis.

[0059] Optionally, in this embodiment, the second unit quantity may be determined from a perspective including, but not limited to, heat dissipation requirements.

[0060] Through the above content, a fully interconnected network architecture is realized in which a first processor with a target bandwidth is connected, ensuring efficient data exchange between computing nodes.

[0061] As an optional implementation, the second connection relationship includes: a second ratio and a third number of second exchange connectors, the second ratio is used to indicate the proportional relationship between the second pin set and the second exchange connector and the proportional relationship between the second exchange connector and the second pin set; the second ratio is determined based on the first connection relationship, the target bandwidth and the bandwidth difference; the third number is calculated based on the first number of computing connectors and the second ratio.

[0062] Optionally, in this embodiment, the third number of second switching connectors is the number of all second switching connectors in the server cabinet.

[0063] Optionally, in this embodiment, the second ratio is used to indicate the proportional relationship between the second pin set and the second switching connector and the proportional relationship between the second switching connector and the second pin set. For example, when the second ratio is 2:1, it means that one second switching connector has a connection relationship with two second pin sets, and one second pin set has a connection relationship with one second switching connector.

[0064] As an optional embodiment, the server cabinet also includes a second switching node chassis, and the second switching connector is deployed in the second switching node chassis; the number of second switching node chassis is the ratio of the third number to the first unit number, and the first unit number is the number of first switching connectors that each first switching node chassis is allowed to accommodate.

[0065] Optionally, in this embodiment, the second switch connectors may be, but are not limited to, evenly distributed on each second switch node chassis.

[0066] Optionally, in this embodiment, the first exchange connector and the second exchange connector may be, but are not limited to, connectors of the same specification attributes. Therefore, when calculating the number of second exchange nodes, it is sufficient to calculate the ratio of the third number to the first unit number corresponding to the attributes of the first exchange connector.

[0067] Through the above content, by clarifying that the number of second switching node chassis is the ratio of the third number (the number of second switching connectors) to the first unit number (the number of connectors that each switching node chassis can accommodate), the rationalization of the number of switching node chassis is ensured, and signal transmission failures caused by the switching node chassis accommodating too many connectors are avoided.

[0068] As an optional implementation, the number of computing connectors is a first number, wherein the first number is determined based on the number of first processors, the server cabinet further includes a computing node chassis, and the computing connectors are deployed in the computing node chassis; the number of computing node chassis is a ratio of the first number to a second unit number, the second unit number being the number of computing connectors that each computing node chassis is allowed to accommodate; and each second switching node chassis is connected to all the computing node chassis under a second connection relationship.

[0069] Optionally, in this embodiment, the number of computing connectors may be, but is not limited to, determined according to the number of first processors. For example, the number of computing connectors may be the number of first processors.

[0070] Optionally, in this embodiment, the computing connectors may be, but are not limited to, evenly distributed on each computing node chassis.

[0071] Optionally, in this embodiment, the second unit quantity may be determined from a perspective including, but not limited to, heat dissipation requirements.

[0072] Through the above content, a fully interconnected connection of signals corresponding to the bandwidth difference is achieved. Combined with the connection between the first pin set having a first connection relationship and the first switching connector, a fully interconnected network architecture is achieved when connecting to a second processor with a reference bandwidth, thereby ensuring efficient data exchange between computing nodes.

[0073] As an optional embodiment, the server cabinet may include but is not limited to: a host computer and a display device, the host computer is connected to the display device, wherein the host computer is used to: obtain processor information and reference bandwidth; match a first connection relationship between a first pin set and a first switching connector according to the processor information; match a second connection relationship between a second pin set and a second switching connector according to a bandwidth difference and the first connection relationship; the display device is used to: display a connector layout diagram of the server cabinet, wherein the connector layout diagram displays a computing connector, a first switching connector, a second switching connector, a first connection relationship, and a second connection relationship that divide the first pin set and the second pin set.

[0074] Optionally, in this embodiment, the host computer can be, but is not limited to, a computer device at a higher level in the control system, which is responsible for receiving data information from the lower computer (such as field equipment or sensors), processing, analyzing and making decisions, and sending control instructions to the lower computer. In the server cabinet, the host computer plays a core role in network management and resource allocation. The host computer can be, but is not limited to, a high-performance server, or a high-performance computer, or a cloud computing platform, or an embedded control unit integrated in the server cabinet, and this application does not limit this.

[0075] Optionally, in this embodiment, the display device may be, but is not limited to, an industrial display or a touch screen operation panel connected to the host computer. The display device may also be, but is not limited to, a mobile device such as a tablet computer or a smart phone, and the display device may communicate with the host computer via a wireless network, but is not limited to.

[0076] Optionally, in this embodiment, the first connection relationship may include, but is not limited to, a first number of calculation connectors, a second number of first exchange connectors, and a first ratio.

[0077] Optionally, in this embodiment, matching the first connection relationship between the first pin set and the first switching connector according to the processor information includes but is not limited to determining a first ratio according to the target bandwidth of the first processor and the switch bandwidth of the switching processor to which the first switching connector is to be connected; determining the second number of the first switching connector according to the first ratio and the first number of computing connectors, wherein the first number of computing connectors is determined according to the number of the first processor. More specifically, determining the first ratio according to the target bandwidth of the first processor and the switch bandwidth of the switching processor to which the first switching connector is to be connected includes but is not limited to matching the first ratio for the available bandwidth combination from bandwidth combinations and ratios having a corresponding relationship, wherein the bandwidth combination is a combination of the bandwidth of the server processor and the bandwidth of the switching processor, and the available bandwidth combination includes the target bandwidth and the switch bandwidth. The corresponding relationship between the bandwidth combination and the ratio can be, but is not limited to, summarized based on historical data.

[0078] Optionally, in this embodiment, the second connection relationship may include, but is not limited to, a third number of second exchange connectors and a second ratio.

[0079] Optionally, in this embodiment, matching the second connection relationship between the second pin set and the second switching connector according to the bandwidth difference and the first connection relationship includes but is not limited to calculating the ratio between the target bandwidth and the bandwidth difference to obtain an intermediate ratio; calculating the second ratio according to the intermediate ratio and the first ratio; and calculating the third number of the second switching connector according to the intermediate ratio and the second number of the first switching connector. For example, when the target bandwidth is 3x, the bandwidth difference is 1x, the first ratio is 2:1, and the second number is y, it can be determined that the second ratio is 6:1 and the third number is y / 3.

[0080] Optionally, in this embodiment, after the host computer determines the first connection relationship and the second connection relationship, the host computer can, but is not limited to, generate a connector layout diagram, and the host computer sends the generated connector layout diagram to a display device, and the display device displays the connector layout diagram after receiving the connector layout diagram sent by the host computer.

[0081] Optionally, in this embodiment, Figure 8 is a connector layout diagram according to an embodiment of the present application. Figure 8As shown, the connector layout diagram may, but is not limited to, show 96 computing connectors, 48 ​​first exchange connectors, 16 second exchange connectors, the first ratio included in the first connection relationship is 2:1, and the second ratio included in the second connection relationship is 6:1. In the connector layout diagram, computing connector parts with the same number have a connection relationship with the first exchange connector part or the second exchange connector part with the same number. For example, the computing connector part numbered 0111 has a connection relationship with the first exchange connector part numbered 0111, and the computing connector part numbered 0112 has a connection relationship with the second exchange connector part numbered 0112.

[0082] Optionally, in this embodiment, the connector layout diagram may be, but is not limited to, matched with the deployment information of the server cabinet, wherein the deployment information of the server cabinet is used to indicate the positional relationship between the computing connector and the first switching connector (or the combination of the first switching connector and the second switching connector) in the server cabinet, and the positional relationship may be, but is not limited to, including the first switching connector (or the combination of the first switching connector and the second switching connector) being located in the middle of the computing connector or the first switching connector (or the combination of the first switching connector and the second switching connector) being located on one side of the computing connector. Figure 8 What is shown is a connector layout diagram in the case where the deployment information is used to indicate that the first switching connector (or the combination of the first switching connector and the second switching connector) in the server cabinet is located in the middle of the computing connector.

[0083] Optionally, in this embodiment, the display device may, but is not limited to, instruct the operator to connect the first pin set and the first switching connector according to a first connection relationship, and to connect the second pin set and the second switching connector according to a second connection relationship through the display of the connector layout diagram.

[0084] Optionally, in this embodiment, the server cabinet may include but is not limited to further including a hardware control module, and the hardware control module may be but is not limited to being connected to a host computer, wherein the hardware control module may be but is not limited to being used to connect a first pin set with a first switching connector according to a first connection relationship, and to connect a second pin set with a second switching connector according to a second connection relationship.

[0085] As an optional implementation, the host computer can be but is not limited to being used to: determine the first number of computing connectors based on the number of first processors; calculate the second number of first switching connectors based on the first number, the target bandwidth, the switch bandwidth of the switching processor to which the first switching connector is to be connected, and a first ratio, wherein the processor information includes: the number of first processors and the target bandwidth, the first connection relationship includes: the first number, the second number and the first ratio, and the first ratio is used to indicate the proportional relationship of the first pin set connecting to the first switching connector and the proportional relationship of the first switching connector connecting to the first pin set.

[0086] Optionally, in this embodiment, the first number may be but is not limited to being equal to the number of first processors.

[0087] As an optional implementation, the host computer can be but is not limited to being used to: calculate the initial number of the first switching connector based on the first number, the target bandwidth, the switch bandwidth and the initial ratio; obtain the first ratio based on the adjustment information of the initial ratio; adjust the initial number according to the first ratio to obtain the second number.

[0088] Optionally, in this embodiment, the initial ratio may be but is not limited to 1:1, that is, each first switching connector is connected to a computing connector (first pin set), and each computing connector (first pin set) is connected to a first switching connector.

[0089] Optionally, it is possible but not limited to that when the initial ratio is used, one first switching connector is connected to one switching processor. Therefore, the initial number of first switching connectors can be obtained by calculating the product of the first number and the target bandwidth and dividing the value of the switch bandwidth.

[0090] Optionally, in this embodiment, it is possible but not limited to that when the initial quantity is determined, the ratio of the initial quantity to the first unit quantity is calculated to determine the initial number of switch node chassis, wherein the first unit quantity is the number of first switch connectors that each first switch node chassis is allowed to accommodate; the initial number of switch node chassis is compared with the first switch node chassis threshold; when the initial number of switch node chassis is greater than the first switch node chassis threshold, the adjustment information is determined to be to expand the initial ratio; when the initial number of switch node chassis is less than or equal to the first switch node chassis threshold, the adjustment information is determined to indicate that the initial ratio has not been adjusted. In the above manner, the situation where the number of required switch node chassis is too large and the server cabinet cannot accommodate it is avoided.

[0091] Optionally, in this embodiment, expanding the initial ratio means expanding the number of computing connectors (first pin sets) connected to a first switching connector. For example, the initial ratio of one first switching connector connected to one computing connector (first pin set) is expanded to one first switching connector connected to two computing connectors (first pin sets).

[0092] Optionally, in this embodiment, when expanding the initial ratio, the number of first switching connectors / second switching connectors that can be accommodated in a switching node chassis and the need to ensure the establishment of a fully interconnected architecture should also be considered.

[0093] Optionally, in this embodiment, the adjustment information may be, but is not limited to, determined based on a ratio adjustment model. The ratio adjustment model is a more sophisticated model that takes into account more factors, such as actual data transmission requirements between computing nodes and switching nodes, system performance bottlenecks, energy efficiency, etc. The system learns and predicts the best first ratio from historical data based on specific information of the first processor, such as processor utilization, task type, etc., to improve signal transmission efficiency and resource utilization.

[0094] Through the above, by introducing the adjustment of the initial ratio, more refined resource management is achieved. Different from the fixed ratio design, this method can dynamically adjust the first ratio and the number of first switching connectors according to the number of first processors and target bandwidth on the current computing node, ensuring a more accurate match between computing resources and network resources, avoiding waste of resources, and improving the efficiency of resource utilization.

[0095] As an optional implementation, the host computer can be but is not limited to being used for: calculating the product of the first quantity and the target bandwidth to obtain a first value when the initial ratio is used to indicate that each computing connector is connected to a first switching connector and each first switching connector is connected to a computing connector; calculating the ratio of the first value to the switch bandwidth to obtain the initial quantity.

[0096] Optionally, in this embodiment, when the initial ratio is used, a first switching connector is connected to a switching processor, and the bandwidth of the first switching connector is the switch bandwidth.

[0097] As an optional implementation, the host computer can be but is not limited to being used for: when the adjustment information is used to indicate that the initial ratio is adjusted and the adjustment method of the initial ratio, adjusting the initial ratio to the first ratio according to the adjustment method; when the adjustment information is used to indicate that the initial ratio is not adjusted, determining the initial ratio as the first ratio.

[0098] Optionally, in this embodiment, when it is detected that the initial number of first switching connectors is greater than or equal to the first switching connector number threshold, the adjustment method of the initial ratio can be obtained, but not limited to, wherein the first switching connector number threshold can be used to indicate the maximum number of first switching connectors that the server cabinet can accommodate. The product of the first unit number and the first switching node number threshold can be determined as the first switching connector number threshold, but not limited to, wherein the first switching node number threshold can be used to indicate the maximum number of first switching node chassis that the server cabinet can accommodate. The server cabinet can include, but not limited to, an input device, which is connected to a host computer, and the host computer can obtain the adjustment method of the initial ratio in the following manner: controlling the display device of the server cabinet to display a ratio adjustment prompt, wherein the ratio adjustment prompt is used to indicate the operator to input the adjustment method of the initial ratio; receiving the adjustment method of the initial ratio input by the operator through the input device.

[0099] Through the above content, choose whether to adjust the initial ratio based on actual conditions. If the initial ratio does not need to be adjusted, select the initial ratio as the first ratio, while ensuring a flexible topology structure and ensuring that the fully interconnected architecture is always established when connecting to the first processor.

[0100] As an optional implementation, the host computer can be but is not limited to being used to: determine a second ratio based on the first connection relationship, the target bandwidth and the bandwidth difference, wherein the second ratio is used to indicate the proportional relationship of the second pin set connecting to the second switching connector and the proportional relationship of the second switching connector connecting to the second pin set; calculate the third number of the second switching connectors based on the first number of calculated connectors and the second ratio, wherein the second connection relationship includes: the second ratio and the third number.

[0101] Through the above content, by determining the second ratio according to the bandwidth difference and the first connection relationship, it is possible to ensure that the server cabinet can flexibly adapt its network connection and data transmission when dealing with processors with different bandwidths. This dynamic adjustment capability enables the server system to effectively utilize existing network resources when dealing with small-bandwidth GPUs, and when dealing with large-bandwidth GPUs, the reasonable allocation of the second switching connector ensures that the additional bandwidth requirements are met, thereby improving the server cabinet's ability to adapt to changing computing needs.

[0102] As an optional implementation, the host computer may be, but is not limited to, further used to: determine the connector bandwidth of the first switching connector according to the first connection relationship and the target bandwidth; and determine the second matching ratio according to the connector bandwidth and the bandwidth difference.

[0103] Optionally, in this embodiment, when the first ratio is not the initial ratio, due to the requirement of connection bandwidth, the number of switching processors connected to a first switching connector may change and may no longer be one. In this case, the connector bandwidth of the first switching connector may not be the switch bandwidth of the switching processor. Therefore, it is necessary to determine the connector bandwidth of the first switching connector based on the first connection relationship and the target bandwidth.

[0104] Optionally, in this embodiment, the second matching ratio is determined according to the connector bandwidth and the bandwidth difference, that is, how many signals with the bandwidth difference can be accommodated by the connector bandwidth of a second switching connector.

[0105] Optionally, in this embodiment, the second ratio may be determined according to the connector bandwidth and the bandwidth difference in the following manner, but is not limited to: a ratio of the connector bandwidth of the first switching connector to the bandwidth difference is calculated to obtain the second ratio.

[0106] As an optional implementation, the host computer may be but is not limited to being further used to: extract the set number of first pin sets connected to each first switching connector from the first connection relationship; and determine the product of the set number and the target bandwidth as the connector bandwidth.

[0107] Optionally, in this embodiment, extracting the number of first pin sets connected to each first switching connector from the first connection relationship includes but is not limited to extracting a proportional relationship between the first switching connectors connecting the first pin sets from the first ratio.

[0108] Through the above content, the connector bandwidth is determined according to the first connection relationship, which ensures the accuracy of the connector bandwidth value and ensures the accurate establishment of the second connection relationship to a certain extent.

[0109] As an optional implementation, the server cabinet also includes a computing node chassis and a first switching node chassis, the computing connector is deployed on the computing node chassis, the first switching connector is deployed on the first switching node chassis, the computing node chassis is used to install the computing node, the first switching node chassis is used to install the first switching node, and the host computer is also used to: calculate the ratio of the second number of first switching connectors to the first unit number to obtain the number of first switching node chassis, wherein the first unit number is the number of first switching connectors that each first switching node chassis is allowed to accommodate, the ratio of the first number to the second unit number is the number of computing node chassis, and the second unit number is the number of computing connectors that each computing node chassis is allowed to accommodate.

[0110] Optionally, in this embodiment, the computing connectors in the server cabinet can be but are not limited to being evenly deployed on the computing node chassis of the server cabinet, and the first switching connector in the server cabinet can be but are not limited to being evenly deployed on the first switching node chassis of the server cabinet.

[0111] As an optional implementation, the display device is also used to: divide the computing connectors according to the number of computing node chassis on the connector layout diagram, and divide the first switching connectors according to the number of first switching node chassis, wherein each first switching node chassis is connected to all the computing node chassis under a first connection relationship.

[0112] Through the above visual division method, the node network architecture design in the server cabinet is more intuitive. When planning or adjusting the node network layout, the operator can intuitively see the connection relationship between the computing node and the first switching node, avoiding complex network calculations and simulations and simplifying the design process.

[0113] As an optional implementation, the server cabinet also includes a computing node chassis and a second switching node chassis, the computing connector is deployed on the computing node chassis, the second switching connector is deployed on the second switching node chassis, the computing node chassis is used to install the computing node, the second switching node chassis is used to install the second switching node, and the host computer is also used to calculate the ratio of the third number of the second switching connectors to the first unit number to obtain the number of the second switching node chassis.

[0114] Optionally, in this embodiment, the second switching connectors in the server cabinet may be, but are not limited to, evenly deployed on the second switching node chassis of the server cabinet.

[0115] Optionally, in this embodiment, the deployment scheme of the second switching connector in the server cabinet on the second switching node chassis may be, but is not limited to, the same as the deployment scheme of the first switching connector in the server cabinet on the first switching node chassis, that is, the number of first switching connectors / second switching connectors deployed on the first switching node chassis and the second switch chassis in the server cabinet may be, but is not limited to, equal, and the interconnection signal resources possessed by the first switching connector deployed on the first switching node chassis in the server cabinet may be, but is not limited to, equal to the interconnection signal resources possessed by the second switching connector deployed on the second switching node chassis in the server cabinet.

[0116] As an optional implementation, the display device is further used to: divide the second switching connectors according to the number of second switching node chassis on the connector layout diagram, wherein each second switching node chassis is connected to all computing node chassis under the second connection relationship.

[0117] Optionally, in this embodiment, a fully interconnected architecture of multiple computing nodes can be realized through, but not limited to, the relay of the first switching node and / or the second switching node. A fully interconnected structure means that a switching node must have a connection relationship with all computing nodes. Therefore, it can be, but not limited to, connecting the first pin set and the first switching connector according to the first connection relationship in the connector layout diagram, and predicting whether the fully interconnected architecture of multiple computing nodes is established after connecting the second pin set and the second switching connector according to the preliminary connection method of the second connection relationship. If it is not established, adjust the preliminary connection method between the second pin set and the second switching connector according to the second connection relationship to obtain the final connection method; connect the second pin set and the second switching connector according to the final connection method in the connector layout diagram.

[0118] An embodiment of the present application also provides a server system. Fig. 9 Schematic diagram of a server system according to an embodiment of the present application. Fig. 9 As shown, the server system includes: a computing node, a switching node and the aforementioned server cabinet, the computing connector included in the server cabinet is connected to the server processor in the computing node, and the switching connector included in the server cabinet is connected to the switching processor in the switching node.

[0119] Optionally, in this embodiment, the server processor may include, but is not limited to, a first processor and a second processor, and the server processor may be evenly deployed on multiple computing nodes, but is not limited to being deployed evenly on multiple computing nodes.

[0120] Optionally, in this embodiment, the exchange connector may include but is not limited to a first exchange connector and a second exchange connector.

[0121] Optionally, in this embodiment, the switching node may include, but is not limited to, a first switching node and a second switching node.

[0122] Optionally, in this embodiment, the switching processors may be, but are not limited to, evenly deployed on one or more switching nodes.

[0123] Optionally, in this embodiment, the server processor may be, but is not limited to, an electronic circuit or device capable of executing instructions, such as a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), an ASIC (Application-Specific Integrated Circuit), or an FPGA (Field-Programmable Gate Array).

[0124] Optionally, in this embodiment, the switching processor may be, but is not limited to, a device responsible for forwarding, processing, and management of data packets, such as a network switching chip or a routing processor.

[0125] Optionally, in this embodiment, each switching node (including the first switching node and the second switching node) can be interconnected but is not limited to being interconnected, and the aforementioned solution realizes that each first switching node is connected to all computing nodes under the first connection relationship and each second switching node is also connected to all computing nodes under the second connection relationship. This makes all computing nodes fully interconnected in the entire server cabinet, so that in the event of a partial switching node failure, each computing node can still communicate through other switching nodes.

[0126] As an optional implementation, the present application also provides a method for non-fixed ratio matching of the cable tray of the whole cabinet server, which solves the problem in the related technology that when designing different GPUs (i.e., server processors) to share a set of computing nodes and cable tray systems, the bandwidth of different GPUs may vary greatly, which may affect the connection topology design of the system.

[0127] If the cable tray system is designed based on the GPU with the largest bandwidth (i.e., the second processor) for GPUs with different design bandwidths, then when installing a small bandwidth GPU later, it is necessary to reduce the connectors (i.e., computing connectors) and cables of the computing nodes (i.e., computing node chassis) based on the existing design and remove the corresponding switching nodes; or only reduce the cables and remove the corresponding switching nodes. The disadvantage of this design is that when the connection relationship ratio (i.e., the first ratio and the second ratio) of the connectors of the computing nodes and the connectors of the switching nodes (i.e., the first switching node chassis and the second switching node chassis) exceeds 1:1, the ratio on each connector is fixed, and the number of corresponding connectors changes with the change of the ratio. For example, when the ratio changes from 1:1 to k:k ratio (k is greater than 1), the number of connectors of the corresponding computing nodes will also change from the original M to M / k; the number of connectors of the switching nodes will also change from the original N to N / k. If M / k or N / k is not an integer, it will affect the existing design.

[0128] The present embodiment adopts a method in which the connection relationship between the connector of the computing node and the connector of the switching node is matched in different ratios when a small-bandwidth GPU (i.e., the first processor) and a large-bandwidth GPU (i.e., the second processor) are installed. This method can be called a Cable Tray non-fixed ratio method. Compared with the method of adopting the same ratio, the design of this method is much more flexible.

[0129] Optionally, in this embodiment, including but not limited to first designing a computing connector within the computing node, all signals of the small-bandwidth GPU and the large-bandwidth GPU are included in the same computing connector, wherein the common signals of the small-bandwidth GPU and the large-bandwidth GPU (i.e., the signals of the target bandwidth) occupy part of the PINs of the computing connector, and the additional signals of the large-bandwidth GPU (i.e., the signals of the bandwidth difference) occupy other PINs of the computing connector.

[0130] Optionally, in this embodiment, including but not limited to, after determining the design scheme of the computing connector, first design the number of GPUs with the minimum bandwidth (i.e., the number of first processors), bandwidth (i.e., target bandwidth), etc., determine the connector conditions of the corresponding switching nodes, etc. At this time, the design of the ratio of the connector of the computing node to the connector of the switching node is formed only according to the actual situation of the small-bandwidth GPU, and the situation of the large-bandwidth GPU is not considered at this time.

[0131] Optionally, in this embodiment, the bandwidth of the small-bandwidth GPU (i.e., the target bandwidth) is set to A, the number of GPUs in a single computing node is N, and the number of computing nodes is P; the bandwidth of the switching chip (i.e., the bandwidth of the switch) is set to B, the number of switching chips in a single switching node is M, and the number of switching nodes is Q. When the ratio of the connector of the computing node to the connector of the switching node is designed to be 1:1 (i.e., the initial ratio), the total bandwidth of the computing node = A × N × P = the total bandwidth of the switching node = B × M × Q. Since the bandwidth of a single small GPU is fixed, the number of GPUs and the number of computing nodes in a single computing node are also fixed according to the original demand, the bandwidth of the switching chip is also fixed, and the number of switching chips in a single switching node needs to be fixed according to the actual heat dissipation, power supply, etc., so it can be basically determined, so the number of switching nodes can also be determined accordingly.

[0132] Optionally, in this embodiment, Fig.10 Schematic diagram of the ratio between the bandwidth of a large-bandwidth GPU and the bandwidth of a small-bandwidth GPU according to an embodiment of the present application. The method provided in the present application can be applied to, but is not limited to, Fig.10 In the GPU scenario shown.

[0133] The connection relationship between the connector and the first switching connector can be calculated according to the initial ratio design but is not limited to it. Under the initial design, a connector of a single computing node is only interconnected with a connector of a single switching node (also called a switch), and a connector of a single switching node is only interconnected with a connector of a single computing node. Therefore, under the original design, the connection relationship between the connector of the computing node and the connector of the switching node is a 1:1 ratio.

[0134] In actual use, if a 1:1 ratio is used, the number of connectors in the computing node / the number of connectors in the switching node = Q / P. Under a 1:1 ratio, the number of connectors in the computing node or the switching node may be too large, and a single computing node or switching node cannot accommodate such a large number of connectors. Therefore, the ratio will be adjusted in actual use to reduce the number of connectors.

[0135] Fig.11 A schematic diagram of a connection relationship between a computing node and a switching node according to an embodiment of the present application Figure 1 .like Fig.11 As shown, when the connection method is designed according to the small bandwidth GPU, there are a total of 16 computing nodes (i.e., computing node 1-computing node 16), and the number of GPUs (i.e., the total number of the first processor / second processor / server processor) is fixed. Then, based on the switching chip bandwidth (i.e., the switch bandwidth of the switching processor), after evaluating the comprehensive design, the number of switching chips (i.e., switching processors) and switches (i.e., switching nodes) can be determined. After that, the number of basic connectors (i.e., computing connectors and first switching connectors) between the computing nodes and switches can be determined (at this time, the ratio is 1:1). Then, according to the actual situation, make adjustments to the ratio and finally determine the final number of connectors for the computing nodes and switches. As shown Fig.11 As shown in , when the computing node uses a small-bandwidth GPU (i.e., the first processor), the interconnection ratio of the connector of the computing node (i.e., the computing connector) and the connector of the switching node (i.e., the first switching connector) is 2:1 (i.e., the first ratio), that is, one connector of each computing node corresponds to one connector of a switching node, and one connector of each switching node corresponds to one connector of each of the two computing nodes. Fig.11 Each computing node shown includes 6 computing connectors, a total of 16 computing nodes. After applying the method provided in the present application, 6 switching nodes are required to achieve full interconnection of computing nodes, and each switching node includes 8 first switching connectors. The connector of each computing node is divided into a first pin set (i.e., a portion numbered XXX1) and a second pin set (i.e., a portion numbered XXX2). The first pin set corresponds to a signal of a small-bandwidth GPU bandwidth, and the second pin set corresponds to a signal of an additional bandwidth of a large-bandwidth GPU compared to a small-bandwidth GPU. Fig.11 There is a connection relationship between the computing connector part and the first switching connector part with the same number in the computing node, for example, the computing connector part in the computing node numbered 0111 and the first switching connector part in the switching node numbered 0111 have a connection relationship.

[0136] After determining the connection relationship of the small-bandwidth GPU part (i.e., the first connection relationship), the second connection relationship is determined based on the first connection relationship and the small GPU bandwidth and the additional bandwidth (the bandwidth difference between the large GPU bandwidth and the small GPU bandwidth). Fig.11 In the case where the bandwidth of the small-bandwidth GPU is 3 / 4 of that of the large-bandwidth GPU, the connector of the computing node should be divided into 4 parts, of which the bandwidth shared by the small-bandwidth GPU and the large-bandwidth GPU occupies 3 / 4 of the connector, and the additional bandwidth of the large-bandwidth GPU occupies the other 1 / 4. The ratio of the small-bandwidth GPU to the switch connector is 2:1 (i.e. the first ratio), that is, the bandwidth of the connectors of every two computing nodes × 3 / 4 of the computing node connector = the bandwidth of the connector of each switching node, plus the bandwidth of the small GPU: the additional bandwidth = 3:1, so it can be learned that the ratio of the large-bandwidth additional part to the switch is 6:1 (i.e. the second ratio), that is, the connectors of every 6 computing nodes (the additional bandwidth part of the large-bandwidth GPU) are connected to 1 connector of the switch.

[0137] Then calculate the additional bandwidth of all 16 computing nodes when installing the high-bandwidth GPU to obtain the number of additional switches required. 16 computing nodes × 6 connectors per computing node / the ratio of the additional bandwidth of the computing node to the switch connector, where each 6 computing nodes' connectors are connected to 1 switch connector = the number of switch connectors required = 16 switch connectors (i.e., the third number of the second switch connector). Divided by the number of connectors per switch (i.e., the first unit number), we get that 2 additional switches are required, and the ratio is 6:1. Therefore, the additional part of the high-bandwidth GPU requires 2 additional switches in the cabinet to support it.

[0138] Since each computing node has exactly 6 connectors, the 6 connectors of each computing node can be used to connect an additional switch node. However, in this case, each additional switch corresponds to 8 computing nodes and cannot be connected to all 16 computing nodes. This architecture is not a fully interconnected architecture.

[0139] This involves considering the specific connection method from the perspective of full interconnection after determining the second ratio and the third number of the second switching connectors. Because the additional bandwidth of the large-bandwidth GPU of the computing node is often less than the shared bandwidth of the large and small GPUs. Therefore, when dealing with the additional bandwidth part, the ratio of the connectors of the computing node to the connectors of the switching node is often large (that is, the connectors of many computing nodes correspond to the connectors of one switching node, which may be equal to or more than the number of connectors owned by each computing node). In this case, it is necessary to ensure the implementation of the fully interconnected architecture. Therefore, when designing the connection topology of the additional bandwidth part according to the 6:1 ratio, the connectors of the selected computing nodes need to be considered. The principle is to make each additional switching node have a connection relationship with each computing node.

[0140] Therefore, the extra part (i.e., the second pin set) and the connector of the switch (i.e., the second switching connector) can be connected by connecting the extra part of the 3 connectors of each computing node + the extra part of the 3 connectors of the next computing node to a node of a switch. In this way, each extra switch can be fully interconnected with 16 computing nodes. Fig.12 A schematic diagram of a connection relationship between a computing node and a switching node according to an embodiment of the present application Figure 2 , Fig.12 The connector parts of the computing nodes and the connector parts of the switching nodes with the same number are connected. Referring to the above design, the following can be finally obtained: Fig.12 The topological connection relationship shown.

[0141] Through the above cable tray non-fixed ratio method, when GPUs with different bandwidths are installed on the computing node, the ratio of the shared PIN (pin) part of the computing node connector and the switch connector is different from the ratio of the extra bandwidth part of the computing node connector and the switch connector. This allows flexible compatibility design based on different GPU bandwidths without affecting the existing topology design.

[0142] 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.

[0143] Specific examples are used herein to illustrate the principles and implementation methods of the present application, and the description of the above embodiments is only used to help understand the method and core ideas of the present application. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present application, several improvements and modifications can be made to the present application, and these improvements and modifications also fall within the scope of protection of the claims of the present application.

Claims

1. A server cabinet, characterized in that: include: a computing connector, a first switching connector, and a second switching connector, wherein the computing connector is divided into a first set of pins and a second set of pins; There is a first connection relationship between the first pin set and the first switching connector, the first connection relationship matches the processor information of the first processor with the target bandwidth, wherein the first pin set corresponds to a signal of the target bandwidth; There is a second connection relationship between the second pin set and the second switching connector, and the second connection relationship matches the bandwidth difference and the first connection relationship, wherein the bandwidth difference is the difference between the reference bandwidth of the second processor and the target bandwidth, and the second pin set corresponds to the signal of the bandwidth difference.

2. The server cabinet according to claim 1, characterized in that: The first connection relationship includes: a first number of the computing connectors, a second number of the first switching connectors, and a first ratio, wherein the first ratio is used to indicate a ratio relationship between the first pin set and the first switching connector and a ratio relationship between the first switching connector and the first pin set; The second number is calculated according to the first number, the target bandwidth, the switch bandwidth of the switch processor to which the first switch connector is to be connected, and the first ratio.

3. The server cabinet according to claim 2, characterized in that: The server cabinet further includes a first switching node chassis, and the first switching connector is disposed on the first switching node chassis; The number of the first switch node chassis is a ratio of the second number to a first unit number, where the first unit number is the number of the first switch connectors that each of the first switch node chassis is allowed to accommodate.

4. The server cabinet according to claim 3, characterized in that: The number of the computing connectors is a first number, wherein the first number is determined according to the number of the first processors, the server cabinet further includes a computing node chassis, and the computing connector is deployed on the computing node chassis; The number of computing node chassis is a ratio of the first number to a second unit number, where the second unit number is the number of computing connectors that each computing node chassis is allowed to accommodate; Each of the first switching node chassis is connected to all of the computing node chassis in the first connection relationship.

5. The server cabinet according to claim 1, characterized in that: The second connection relationship includes: a second ratio and a third number of the second switching connectors, the second ratio being used to indicate a ratio relationship between the second pin set and the second switching connector and a ratio relationship between the second switching connector and the second pin set; The second matching ratio is determined according to the first connection relationship, the target bandwidth and the bandwidth difference; The third quantity is calculated based on the first quantity of the calculation connectors and the second ratio.

6. The server cabinet according to claim 5, characterized in that: The server cabinet further includes a second switch node chassis, and the second switch connector is disposed in the second switch node chassis; The number of the second switch node chassis is a ratio of the third number to a first unit number, where the first unit number is the number of the first switch connectors that each first switch node chassis is allowed to accommodate.

7. The server cabinet according to claim 6, characterized in that: The number of the computing connectors is a first number, wherein the first number is determined according to the number of the first processors, the server cabinet further includes a computing node chassis, and the computing connector is deployed in the computing node chassis; The number of computing node chassis is a ratio of the first number to a second unit number, where the second unit number is the number of computing connectors that each computing node chassis is allowed to accommodate; Each of the second switching node chassis is connected to all of the computing node chassis in the second connection relationship.

8. The server cabinet according to claim 1, characterized in that: The server cabinet further includes: a host computer and a display device, wherein the host computer is connected to the display device, The host computer is used to: obtain the processor information and the reference bandwidth; match the first connection relationship between the first pin set and the first switching connector according to the processor information; match the second connection relationship between the second pin set and the second switching connector according to the bandwidth difference and the first connection relationship; The display device is used to: display a connector layout diagram of the server cabinet, wherein the connector layout diagram shows the computing connector that divides the first pin set and the second pin set, the first switching connector, the second switching connector, the first connection relationship, and the second connection relationship.

9. The server cabinet according to claim 8, characterized in that: The host computer is also used for: Determining a first number of the computing connectors according to the number of the first processors; The second quantity of the first switching connectors is calculated based on the first quantity, the target bandwidth, the switch bandwidth of the switching processor to which the first switching connector is to be connected, and a first ratio, wherein the processor information includes: the quantity of the first processors and the target bandwidth, the first connection relationship includes: the first quantity, the second quantity and the first ratio, and the first ratio is used to indicate the proportional relationship of the first pin set connecting to the first switching connector and the proportional relationship of the first switching connector connecting to the first pin set.

10. The server cabinet according to claim 9, characterized in that: The host computer is also used for: Calculating an initial number of the first switching connectors according to the first number, the target bandwidth, the switch bandwidth, and an initial ratio; Acquiring the first ratio according to the adjustment information of the initial ratio; The initial quantity is adjusted according to the first ratio to obtain the second quantity.

11. The server cabinet according to claim 10, characterized in that: The host computer is also used for: In a case where the initial ratio is used to indicate that each computing connector is connected to one first switching connector and each first switching connector is connected to one computing connector, calculating a product of the first number and the target bandwidth to obtain a first value; The ratio of the first value to the switch bandwidth is calculated to obtain the initial number.

12. The server cabinet according to claim 10, characterized in that: The host computer is also used for: In a case where the adjustment information is used to indicate that the initial ratio is adjusted and the adjustment method of the initial ratio, adjusting the initial ratio to the first ratio according to the adjustment method; When the adjustment information indicates that the initial ratio has not been adjusted, the initial ratio is determined as the first ratio.

13. The server cabinet according to claim 8, characterized in that: The host computer is also used for: Determine a second ratio according to the first connection relationship, the target bandwidth, and the bandwidth difference, wherein the second ratio is used to indicate a ratio relationship between the second pin set and the second switching connector and a ratio relationship between the second switching connector and the second pin set; The third quantity of the second exchange connectors is calculated according to the first quantity of the calculated connectors and the second ratio, wherein the second connection relationship includes: the second ratio and the third quantity.

14. The server cabinet according to claim 13, characterized in that: The host computer is also used for: Determine a connector bandwidth of the first switching connector according to the first connection relationship and the target bandwidth; The second matching ratio is determined according to the connector bandwidth and the bandwidth difference.

15. The server cabinet according to claim 14, characterized in that: The host computer is also used for: Extracting the set quantity of the first pin sets connected to each of the first switching connectors from the first connection relationship; The connector bandwidth is determined as a product of the set quantity and the target bandwidth.

16. The server cabinet according to claim 9, characterized in that: The server cabinet further includes a computing node chassis and a first switching node chassis, the computing connector is deployed on the computing node chassis, the first switching connector is deployed on the first switching node chassis, the computing node chassis is used to install a computing node, the first switching node chassis is used to install a first switching node, and the host computer is further used to: Calculate the ratio of the second quantity of the first switching connectors to the first unit quantity to obtain the quantity of the first switching node chassis, wherein the first unit quantity is the quantity of the first switching connectors that each of the first switching node chassis is allowed to accommodate, the ratio of the first quantity to the second unit quantity is the quantity of the computing node chassis, and the second unit quantity is the quantity of the computing connectors that each of the computing node chassis is allowed to accommodate.

17. The server cabinet according to claim 16, characterized in that: The display device is further used for: The computing connectors are divided according to the number of the computing node chassis on the connector layout diagram, and the first switching connectors are divided according to the number of the first switching node chassis, wherein each of the first switching node chassis is connected to all the computing node chassis under the first connection relationship.

18. The server cabinet according to claim 16, characterized in that: The server cabinet further includes a computing node chassis and a second switching node chassis, the computing connector is deployed on the computing node chassis, the second switching connector is deployed on the second switching node chassis, the computing node chassis is used to install a computing node, the second switching node chassis is used to install a second switching node, and the host computer is further used to: The number of the second switch node chassis is obtained by calculating a ratio of a third number of the second switch connectors to the first unit number.

19. The server cabinet according to claim 18, characterized in that: The display device is further used for: The second switch connectors are divided according to the number of the second switch node chassis on the connector layout diagram, wherein each of the second switch node chassis is connected to all of the computing node chassis under the second connection relationship.

20. A server system, characterized in that: include: A computing node, a switching node and a server cabinet as described in any one of claims 1 to 19, wherein the computing connector included in the server cabinet is connected to the server processor in the computing node, and the switching connector included in the server cabinet is connected to the switching processor in the switching node.

Images