GPU cluster system based on optical interconnection

Abstract

The invention provides a GPU cluster system based on optical interconnection, which comprises at least one cluster unit, each cluster unit comprises a control module, an optical communication module and a plurality of computing nodes, the control module and the computing nodes are respectively connected with the optical communication module through optical fibers, and the optical communication module realizes signal transmission between the control module and the computing nodes; the control module is used for distributing calculation tasks in the cluster unit and scheduling calculation nodes; each computing node is configured with at least one graphics processor for executing the assigned computing task using the graphics processor when scheduled. By fully utilizing the characteristics of high speed and low delay of the optical interconnection technology, the communication efficiency among devices, especially among computing nodes, is optimized, the flexible expansion and configuration of GPU chips and CPUs are realized, the computing power scale and performance of the whole intelligent computing cluster are improved, and the increasing computing power demand is met.

Description

Technical Field

[0001] The present invention relates to the field of high performance computing technology, and in particular to a GPU cluster system based on optical interconnection. Background Art

[0002] With the rapid development of technologies such as artificial intelligence and big data, the demand for computing power is growing. As the core component of high-performance computing, GPU (graphics processing unit) plays an important role in accelerating computing tasks. In recent years, domestic GPU chip products have gradually emerged, providing strong support for building an independent and controllable high-performance computing platform. However, the existing GPU chip clusters still have deficiencies in terms of inter-node communication efficiency, computing power expansion and configuration flexibility, making it difficult to meet the growing computing power demand.

[0003] Optical interconnect technology has significant advantages in improving communication efficiency due to its high speed and low latency. However, applying optical interconnect technology to the construction of GPU chip clusters requires solving a series of problems such as technology adaptation, cluster performance testing, supporting equipment and software platform requirements. Summary of the invention

[0004] Based on the above problems, the present invention provides a GPU cluster system based on optical interconnection, aiming to solve the technical problems of low communication efficiency and inflexible configuration of GPU clusters in the prior art.

[0005] A GPU cluster system based on optical interconnection includes at least one cluster unit, each cluster unit includes a control module, an optical communication module and a plurality of computing nodes, the control module and the computing nodes are connected via optical fibers and the optical communication module respectively, and the optical communication module realizes signal transmission between the control module and the computing nodes; the control module is used for allocating computing tasks within the cluster unit and scheduling computing nodes; each computing node is configured with at least one graphics processor for using the graphics processor to execute the allocated computing tasks when being scheduled.

[0006] Furthermore, the control module includes a central processing unit; when the control module includes two or more central processing units, the central processing units are connected via a unified platform interface.

[0007] Furthermore, the optical communication module includes at least two optical switches, each of which is connected to the control module via optical fiber; each optical switch is connected to a group of computing nodes, each computing node within each group of computing nodes is connected to the corresponding optical switch via optical fiber, and the computing nodes within each group of computing nodes are connected to each other via optical fiber for signal transmission.

[0008] Furthermore, each optical switch is configured with a network card. When the cluster system includes two or more cluster units, information is transmitted between different cluster units through the network card.

[0009] Furthermore, it includes two cluster units, and the optical communication module of each cluster unit includes two optical switches; the first optical switch of the first cluster unit and the second optical switch of the second cluster unit transmit information through a network card; the second optical switch of the first cluster unit and the first optical switch of the second cluster unit transmit information through a network card.

[0010] Furthermore, one of the types of computing tasks is a training task of a machine learning model.

[0011] Furthermore, the control module is also used to monitor each computing node in the cluster unit and the graphics processor in the computing node in real time to obtain real-time monitoring results.

[0012] Furthermore, the control module is also used to perform a performance test on the cluster unit to obtain a performance test result.

[0013] Furthermore, the control module is also used to perform BIOS configuration on each computing node in the cluster unit.

[0014] Furthermore, the control module is also used to perform BMC configuration on each computing node in the cluster unit.

[0015] The beneficial technical effect of the present invention is that by making full use of the high-speed, low-latency characteristics of optical interconnection technology, the communication efficiency between devices, especially between computing nodes, is optimized, and the needs of optical interconnection technology adaptation, cluster performance testing, etc. are met. Users can appropriately expand the number of computing nodes and the number of cluster units according to needs, realize flexible expansion and configuration of GPU chips and CPUs, improve the computing power scale and performance of the entire intelligent computing cluster, and meet the growing computing power needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A schematic diagram of a module of a GPU cluster system based on optical interconnection according to the present invention; Figure 2 The figure is a schematic diagram of the network topology structure of a specific implementation mode of a GPU cluster system based on optical interconnection of the present invention. DETAILED DESCRIPTION

[0017] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0018] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0019] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, but they are not intended to limit the present invention.

[0020] See also Figure 1-Figure 2 The present invention provides a GPU cluster system based on optical interconnection, including at least one cluster unit, each cluster unit includes a control module 1, an optical communication module 2 and a plurality of computing nodes 3, the control module 1 and the computing nodes 3 are respectively connected through optical fibers and the optical communication module 2, and the optical communication module 2 realizes signal transmission between the control module 1 and the computing nodes 3; the control module 1 is used to allocate computing tasks within the cluster unit and schedule computing nodes; each computing node 3 is configured with at least one graphics processor, which is used to use the graphics processor to execute the allocated computing tasks when being scheduled.

[0021] Optical interconnection technology uses optical fiber as a transmission medium and realizes high-speed signal transmission through the principle of total reflection of light. Compared with traditional electrical interconnection technology, optical interconnection has the advantages of large bandwidth, low latency, and strong anti-interference ability, making it very suitable for building large-scale parallel computing clusters.

[0022] When there are two or more cluster units, information is transmitted between the cluster units through optical communication modules.

[0023] Specifically, each computing node is a GPU (graphics processing unit) server.

[0024] Specifically, the graphics processor is a high-performance domestic GPU chip type, which is responsible for performing computing tasks.

[0025] As an implementation manner of the present invention, the GPU chip type of each computing node may be different, and the GPU chips in the same computing node may also be different.

[0026] Furthermore, the control module 1 includes a central processing unit; when the control module 1 includes two or more central processing units, the central processing units are connected via a unified platform interface.

[0027] Control module 1 uses the central processing unit (CPU) to perform general calculations and logic processing, schedule computing nodes, and allocate computing tasks, providing support for high concurrency and multi-threaded processing of the cluster.

[0028] The unified platform interface, or UPI (Unified Platform Interface), provides high-performance interconnection between CPUs to form a tightly coupled CPU cluster. It has the characteristics of high bandwidth, low latency, high scalability, flexible topology, and improved power management. It provides faster data transmission speed, lower latency, and better energy efficiency between CPUs, which helps to achieve fast data sharing and collaborative work between CPUs and meet the needs of high-performance computing. By making full use of the high-speed and low-latency characteristics of optical interconnection technology, the communication efficiency between devices, especially between computing nodes, is optimized, and the needs of optical interconnection technology adaptation and cluster performance testing are met. Users can appropriately expand the number of computing nodes and the number of cluster units according to their needs, realize flexible expansion and configuration of GPU chips and CPUs, and improve the computing power scale and performance of the entire intelligent computing cluster to meet the growing computing power needs.

[0029] Furthermore, the optical communication module 2 includes at least two optical switches, each of which is connected to the control module 1 via optical fiber; each optical switch is connected to a group of computing nodes, each computing node 3 within each group of computing nodes is connected to the corresponding optical switch via optical fiber, and the computing nodes 3 within each group of computing nodes are connected to each other via optical fiber for signal transmission.

[0030] Optical switches constitute the core of optical internetworks. Through optical switches and optical fiber cabling, efficient data transmission between GPU chips can be achieved, communication bottlenecks can be reduced, and overall computing performance can be improved. They are very suitable for building large-scale parallel computing clusters.

[0031] Specifically, when the control module has multiple central processors, all optical switches are connected to one of the central processors for data transmission.

[0032] Optical interconnection technology is used to achieve high-speed, low-latency data transmission between GPU servers, thereby improving overall computing efficiency.

[0033] Furthermore, each optical switch is configured with a network card 4. When the cluster system includes two or more cluster units, information is transmitted between different cluster units through the network card.

[0034] The function of the network card is to transfer the computing data within the cluster unit to another cluster unit.

[0035] As a preferred implementation of the present invention, the network card realizes information transmission between different cluster units through optical interconnection network technology.

[0036] Furthermore, it includes two cluster units, and the optical communication module 2 of each cluster unit includes two optical switches; the first optical switch of the first cluster unit and the second optical switch of the second cluster unit transmit information through the network card 4; the second optical switch of the first cluster unit and the first optical switch of the second cluster unit transmit information through the network card 4.

[0037] The present invention forms a complex network topology through the combination of optical interconnection paths (i.e., optical fiber wiring) of cluster units and optical switches. The control module and the optical switch can communicate with each other through optical fiber, the optical switch and the computing node can communicate with each other through optical fiber, the computing nodes in the group can communicate with each other through optical fiber, and the optical switches of different cluster units communicate with each other through network cards. Multiple communication modes and data transmission strategies are supported. This structure enables the various GPU servers in the cluster system to flexibly cooperate with each other to jointly complete complex computing tasks.

[0038] According to the network topology of the present invention, the central processor in the control module can be expanded as needed, the number of optical switches can be expanded as needed, the number of computing nodes connected to each optical switch can be expanded, and even one or more cluster units can be expanded as needed. The network topology of the present invention achieves good scalability. As computing needs grow, the computing power of the cluster can be expanded by adding more GPU servers and / or optical switches. Since optical interconnection technology has the advantages of large bandwidth and low latency, the expanded cluster can still maintain efficient computing performance.

[0039] Furthermore, one of the types of computing tasks is a training task of a machine learning model.

[0040] The network topology architecture of the present invention can greatly improve the efficiency of parallel computing, shorten the training time, and improve resource utilization when conducting large-scale machine learning model training. It can achieve parallel acceleration of large model training under domestic AI computing power clusters, breaking through the traditional electrical interconnection limitations, and significantly improving the data transmission speed and the overall performance of the system. Moreover, the network topology architecture of the present invention is highly versatile and can be combined with different types of GPU chips to provide high-speed, low-energy interconnection between GPU chips.

[0041] Furthermore, the control module 1 is also used to perform real-time monitoring on each computing node 3 in the cluster unit and the graphics processor in the computing node 3 to obtain real-time monitoring results.

[0042] Furthermore, the control module 1 is further used to perform a performance test on the cluster unit to obtain a performance test result. Furthermore, the control module 1 is further used to perform a BIOS configuration on each computing node in the cluster unit.

[0043] Furthermore, the control module 1 is also used to perform BMC configuration on each computing node in the cluster unit.

[0044] BIOS (Basic Input / Output System) and BMC (Baseboard Management Controller) are two different components in a server or computer, each of which has different functions. BIOS is a set of programs fixed in the ROM chip on the computer motherboard. It provides the lowest-level and most direct hardware settings and controls for the computer, and implements functions such as operating system initialization. BMC is a dedicated microcontroller integrated on the server motherboard for out-of-band management and monitoring functions. Customize the BIOS and BMC for each computing node to ensure that the cluster system runs in the optimal performance mode.

[0045] Of course, you can also use other methods to customize the BIOS and BMC configuration of each computing node instead of using the control module.

[0046] The server hardware of the GPU computing node of the present invention includes a motherboard, CPU, memory, disk or SSD, network card, etc., ensuring that the hardware performance meets the cluster operation requirements.

[0047] In addition, the cluster system uses distributed storage to store computing data, improving the reliability and scalability of data storage.

[0048] The above are only preferred embodiments of the present invention, and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should be aware that all solutions obtained by equivalent substitutions and obvious changes made using the description and illustrations of the present invention should be included in the protection scope of the present invention.

Claims

1. A GPU cluster system based on optical interconnection, characterized in that: It comprises at least one cluster unit, each of which comprises a control module, an optical communication module and a plurality of computing nodes, wherein the control module and the computing nodes are respectively connected to the optical communication module via optical fibers, and the optical communication module realizes signal transmission between the control module and the computing nodes; the control module is used for allocating computing tasks within the cluster unit and scheduling the computing nodes; each of the computing nodes is configured with at least one graphics processor for using the graphics processor to execute the allocated computing tasks when being scheduled.

2. The GPU cluster system based on optical interconnection according to claim 1, characterized in that: The control module includes a central processing unit; when the control module includes two or more central processing units, the central processing units are connected via a unified platform interface.

3. The GPU cluster system based on optical interconnection according to claim 1, characterized in that: The optical communication module includes at least two optical switches, each of which is connected to the control module via an optical fiber; each of the optical switches is connected to a group of computing nodes, each of the computing nodes in each group of computing nodes is connected to the corresponding optical switch via an optical fiber, and the computing nodes in each group of computing nodes are connected to each other via optical fibers for signal transmission.

4. The GPU cluster system based on optical interconnection as claimed in claim 3, characterized in that: Each of the optical switches is configured with a network card. When the cluster system includes two or more cluster units, information is transmitted between different cluster units through the network card.

5. The GPU cluster system based on optical interconnection as claimed in claim 4, characterized in that: It comprises two cluster units, and the optical communication module of each cluster unit comprises two optical switches; the first optical switch of the first cluster unit and the second optical switch of the second cluster unit transmit information through a network card; the second optical switch of the first cluster unit and the first optical switch of the second cluster unit transmit information through a network card.

6. The GPU cluster system based on optical interconnection according to claim 1, characterized in that: One of the types of computing tasks is a training task of a machine learning model.

7. The GPU cluster system based on optical interconnection according to claim 1, characterized in that: The control module is also used to monitor each of the computing nodes in the cluster unit and the graphics processor in the computing node in real time to obtain real-time monitoring results.

8. The GPU cluster system based on optical interconnection according to claim 1, characterized in that: The control module is also used to perform a performance test on the cluster unit to obtain a performance test result.

9. The GPU cluster system based on optical interconnection according to claim 1, characterized in that: The control module is also used to perform BIOS configuration on each of the computing nodes in the cluster unit.

10. The GPU cluster system based on optical interconnection according to claim 1, characterized in that: The control module is also used to perform BMC configuration on each computing node in the cluster unit.

Images