CN-122027921-A - Sparse optical communication direct connection system and method

CN122027921ACN 122027921 ACN122027921 ACN 122027921ACN-122027921-A

Abstract

The invention provides a sparse optical communication direct connection system and a sparse optical communication direct connection method. The sparse optical communication direct connection system comprises a plurality of computing nodes, wherein any two computing nodes in the plurality of computing nodes are in optical connection, each computing node comprises a plurality of GPUs, any two GPUs in the same computing node are in electrical connection, any two GPUs in different computing nodes are in optical connection, and the GPUs are used for executing model reasoning computing tasks. In the mode, a layered sparse optical direct-connection architecture is provided, and the contradiction between the full interconnection cost and the performance can be solved through the software and hardware collaborative design.

Inventors

Request for anonymity
ZHANG RUITAO

Assignees

光子算数(南京)科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260212

Claims (10)

1. The sparse optical communication direct connection system is characterized by comprising a plurality of computing nodes, wherein any two of the computing nodes are optically connected; any two GPUs in the same computing node are electrically connected, and any two GPUs in different computing nodes are optically connected; The GPU is used for executing model reasoning calculation tasks.
2. The sparse optical communication direct connection system of claim 1, wherein any two of the GPUs in the same compute node are electrically connected by a PCIe bus and interface of the compute node.
3. The sparse optical communication direct connection system according to claim 1, wherein in each computing node, each GPU is correspondingly provided with an optical interconnection interface, and any two GPUs in different computing nodes are optically connected through the optical interconnection interfaces.
4. The sparse optical communication direct connection system of claim 1, wherein any two of the plurality of computing nodes are optically connected by an optical fiber.
5. The sparse optical communication direct connection system of claim 3, wherein the number of compute nodes is N, and the number of optical interconnect interfaces each compute node comprises is N, N being greater than 1; n-1 optical interconnection interfaces in each computing node are respectively and optically connected with one optical interconnection interface in other N-1 computing nodes through optical fibers.
6. The sparse optical communication direct connection system of claim 5, wherein one of the optical interconnect interfaces of each of the compute nodes that is not optically connected to the other N-1 compute nodes is an idle exit for larger scale cluster training.
7. The sparse optical communication direct connection system of any one of claims 1-6, wherein the sparse optical communication direct connection system comprises 8 compute nodes, each compute node comprising 8 GPUs.
8. A sparse optical communication direct connection method, applied to the sparse optical communication direct connection system of any one of claims 1-7, comprising: Identifying a communication operator type after the distributed training or reasoning task is started; determining a communication route based on the communication operator type; generating a dynamic messaging interface communication plan based on the communication route; invoking a driver to perform the data transmission of the communication plan.
9. The method of claim 8, wherein the step of determining a communication route based on the communication operator type comprises: if the communication operator type is set communication, determining a communication route as a fixed route based on a logic ring; And if the communication operator type is the mixed expert model dynamic route, determining the communication route as a point-to-point direct connection route combined with a real-time load balancing strategy.
10. The method of claim 8, wherein after the step of invoking the drive to perform data transmission of the communication plan, the method further comprises: after the end of the data transmission of the communication plan, the next calculation step of the distributed training or reasoning task is performed.

Description

Sparse optical communication direct connection system and method Technical Field The invention relates to the technical field of high-performance computer interconnection networks, in particular to a sparse optical communication direct connection system and method. Background Traditional electrical interconnection technology (such as Ethernet and InfiniBand (a high-performance and low-delay ‌ computer network communication standard)) depends on that an electronic signal is transmitted through a copper wire during data transmission, the performance of the traditional electrical interconnection technology is inherently limited by physical and electrical characteristics, the frequency of an electrical signal is improved, the signal integrity is deteriorated due to skin effect, crosstalk and other problems, high-frequency bottlenecks (such as 400G and more complex coding and equalization technology) are difficult to break through, long-distance transmission needs relay amplification, the high-frequency signal impedance loss is obvious, the energy consumption is exponentially increased along with the increase of bandwidth, and meanwhile, the additional delay is introduced due to the fact that protocol stack processing (such as TCP/IP (transmission control protocol/Internet protocol ‌, transmission Control Protocol/Internet Protocol) layering) is caused, and microsecond-level requirements of artificial intelligent training and reasoning scenes are difficult to meet. These limitations have prompted new technologies for optical interconnects to become high bandwidth, low power consumption, low latency alternatives. The optical interconnection technology fundamentally breaks through the physical limitation of electrical interconnection through laser transmission signals, namely, single-fiber multichannel parallel transmission is realized through wavelength division multiplexing (WDM, WAVELENGTH DIVISION MULTIPLEXING) by utilizing the high-frequency characteristic of light (such as that a single-mode fiber can reach THz-level bandwidth), TB-level bandwidth is easily supported, the photon transmission almost has no resistance loss, relay amplification is not needed in a long distance, the distance can reach tens to hundreds of meters, the protocol conversion level can be reduced by an optical transmission architecture, and the end-to-end delay can be reduced to nanosecond level. With the continuous expansion of the scale of artificial intelligence models, single machine training and reasoning can not meet the large-scale requirement of large models, so that large-scale calculation power clusters become a necessary choice. In the large model reasoning training process, a close cooperative relationship exists between a communication operator and a computing power cluster topology, wherein the communication operator (such as AllReduce, allGather, alltoAll and the like) is in charge of key data interaction such as gradient synchronization and parameter aggregation among distributed computing nodes, and the efficiency is directly limited by a cluster topology structure (such as high-speed connection among GPU (graphic processor, graphics Processing Unit) cards, a connection protocol among servers, physical layout of hardware facilities, switch layers and the like). The optimized topology design can reduce the communication delay of the cross-node and match operator characteristics (such as a hierarchical aggregation strategy), otherwise, the inefficient topology can cause the communication to become a training bottleneck. Therefore, in actual deployment, topology needs to be dynamically adapted according to the communication mode (data quantity, frequency and dependency relationship) of operators, and the computing power of the clusters is most effectively exerted. Disclosure of Invention In view of the above, the present invention aims to provide a sparse optical communication direct connection system and method, so as to provide a layered sparse optical direct connection architecture, and solve the contradiction between the total interconnection cost and the performance through the software and hardware collaborative design. In a first aspect, an embodiment of the present invention provides a sparse optical communication direct connection system, where the sparse optical communication direct connection system includes a plurality of computing nodes, any two computing nodes in the plurality of computing nodes are optically connected, each computing node includes a plurality of GPUs, any two GPUs in the same computing node are electrically connected, and any two GPUs in different computing nodes are optically connected, and the GPUs are configured to perform model reasoning computing tasks. In an alternative embodiment of the present application, any two GPUs in the same computing node are electrically connected through a PCIe bus and an interface of the computing node. In an optional embodiment of the present application,