Search

CN-122001956-A - Data transmission method and device of server, readable storage medium and server

CN122001956ACN 122001956 ACN122001956 ACN 122001956ACN-122001956-A

Abstract

The application provides a data transmission method and device of a server, a readable storage medium and the server, and relates to the technical field of data transmission. The data transmission method comprises the steps that a first computing chip generates an Ethernet frame based on data to be transmitted, the Ethernet frame comprises an Ethernet grouping head, the Ethernet grouping head comprises an address field, the address field is used for indicating a physical memory address of the first computing chip and a physical memory address of a second computing chip, the second computing chip is a chip for receiving the Ethernet frame, and the Ethernet switching chip sends the Ethernet frame to the second computing chip based on the Ethernet grouping head. The application changes the Ethernet frame header field which is originally invalid overhead into the effective field carrying information related to the effective load, thereby avoiding the problem of space waste caused by the Ethernet packet header originally when the small-scale AI server realizes the mutual transmission of chip data through the Ethernet and improving the efficiency of the data interaction between chips.

Inventors

  • GUO CHUANXIONG

Assignees

  • 北京比特智路信息技术有限公司

Dates

Publication Date
20260508
Application Date
20241106

Claims (12)

  1. 1. A data transmission method of a server, wherein the server includes an ethernet switching chip and at least two computing chips, the data transmission method comprising: the method comprises the steps that a first computing chip generates an Ethernet frame based on data to be transmitted, wherein the Ethernet frame comprises an Ethernet packet header, the Ethernet packet header comprises an address field, the address field is used for indicating a physical memory address of the first computing chip and a physical memory address of a second computing chip, and the second computing chip is a chip for receiving the Ethernet frame; The Ethernet switch chip sends the Ethernet frame to the second computing chip based on the Ethernet packet header.
  2. 2. The method according to claim 1, wherein the address field includes a target address field, the target address field being used to indicate a physical memory address of the second computing chip; Or the address field comprises the target address field and a source address field, wherein the source address field is used for indicating a physical memory address of the first computing chip or identification information of the first computing chip; Wherein the destination address field is located before the source address field, or the source address field is located before the destination address field.
  3. 3. The method according to claim 2, wherein the destination address field includes a chip identification field, the length of the chip identification field being smaller than the length of the destination address field, the chip identification field being used to indicate a chip identification of the second computing chip; the Ethernet switching chip sends the Ethernet frame to the second computing chip based on the chip identification field.
  4. 4. The data transmission method of a server according to claim 1, wherein the ethernet packet header further includes a message type field for indicating a message type of the ethernet frame and a data packet type field for indicating a data packet of the ethernet frame; wherein the message type comprises a read operation, a write operation or an atomic operation, and the data packet comprises a message start, a message end or an intermediate message.
  5. 5. The data transmission method of a server according to claim 1, wherein the ethernet packet header further includes an ethernet type field for indicating a message type of the ethernet frame, a message type field for indicating a data packet of the ethernet frame, and a data packet type field; The message type comprises a read operation, a write operation or an atomic operation, the data packet comprises a message start, a message end or an intermediate message, the Ethernet type field is positioned before the message type field, and the message type field is positioned before the data packet type field.
  6. 6. The data transmission method of the server according to claim 4 or 5, wherein the ethernet packet header further includes a queue sequence number field for indicating a connection identification of the ethernet frame and a packet sequence number field for indicating a packet sequence number of the ethernet frame; wherein the queue sequence number field is located after the data packet type field and the packet sequence number field is located after the queue sequence number field.
  7. 7. The data transmission method of a server according to claim 4 or 5, wherein the ethernet frame further includes an extension header; The extension header is used for indicating the read-write length of the Ethernet frame when the message type is a read operation or a write operation, and is used for indicating the operand of the atomic operation when the message type is an atomic operation.
  8. 8. The data transmission method of a server according to any one of claims 1 to 5, wherein the ethernet frame further includes a payload field and a frame check sequence field; Wherein the payload field is located after the ethernet packet header and the frame check sequence field is located after the payload field.
  9. 9. A data transmission device of a server, wherein the server includes an ethernet switching chip and at least two computing chips, the data transmission device comprising: The system comprises a generation module, a first computing chip, a second computing chip and a data transmission module, wherein the generation module is used for controlling the first computing chip to generate an Ethernet frame based on data to be transmitted, the Ethernet frame comprises an Ethernet packet header, the Ethernet packet header comprises an address field, the address field is used for indicating a physical memory address of the first computing chip and a physical memory address of the second computing chip, and the second computing chip is a chip for receiving the Ethernet frame; and the sending module is used for controlling the Ethernet switching chip to send the Ethernet frame to the second computing chip based on the Ethernet packet header.
  10. 10. A data transmission apparatus of a server, comprising: a memory having a program or instructions stored thereon; a processor for implementing the steps of the data transmission method of a server according to any one of claims 1 to 8 when executing the program or instructions.
  11. 11. A readable storage medium having stored thereon a program or instructions, which when executed by a processor, implement the steps of the data transmission method of a server according to any of claims 1 to 8.
  12. 12. A server is characterized in that, the server includes: A data transmission device of a server according to claim 9 or 10, and/or The readable storage medium of claim 11.

Description

Data transmission method and device of server, readable storage medium and server Technical Field The present application relates to the field of data transmission technologies, and in particular, to a data transmission method and apparatus for a server, a readable storage medium, and a server. Background In the related art, a large number of GPUs (Graphics Processing Unit, graphics processors) or AI (ARTIFICIAL INTELLIGENCE ) chips are required for training artificial intelligence such as a large language model to provide computing power, and a current trend is to integrate a plurality of GPUs or AI chips in one server, so that one server has stronger computing power. Multiple GPU or AI chips in a server need to be interconnected through a high-speed network, so as to realize data transmission, and it is common practice to realize non-blocking interconnection between multiple chips based on an ethernet switch chip. In the ethernet-based chip interconnection Transport protocol packet format, an additional TL Hdr (Transport LAYER HEADER) field is added after the field of the ethernet packet header, where the TL Hdr field includes a memory operation type, a source chip physical memory address, a destination chip physical memory address, and the like. The ethernet packet header carries a packet destination MAC (MEDIA ACCESS Control Address) Address, where the MAC Address is only of local significance, and is used to identify an ethernet card port, and does not carry information related to a payload, and IP (Internet Protocol) routing is not needed in a small-scale ethernet network, so that fields of the ethernet packet header form extra Address overhead, resulting in space waste and low efficiency of data interaction. Disclosure of Invention The present application aims to solve at least one of the technical problems existing in the prior art or related art. To this end, a first aspect of the present application proposes a data transmission method of a server. A second aspect of the present application proposes a data transmission device of a server. A third aspect of the present application proposes a data transmission device of a server. A fourth aspect of the application proposes a readable storage medium. A fifth aspect of the present application proposes a server. In view of this, a first aspect of the present application provides a data transmission method of a server, where the server includes an ethernet switch chip and at least two computing chips, the data transmission method includes that the first computing chip generates an ethernet frame based on data to be transmitted, where the ethernet frame includes an ethernet packet header, the ethernet packet header includes an address field, the address field is used to indicate a physical memory address of the first computing chip and a physical memory address of a second computing chip, the second computing chip is a chip that receives the ethernet frame, and the ethernet switch chip sends the ethernet frame to the second computing chip based on the ethernet packet header. In this technical solution, the server includes an AI server, and the server includes a plurality of computing chips, where the computing chips may be GPU, NPU (Neural Processing Unit, neural network processor), or AI chips. The computing chip can provide computing power for training processes or reasoning processes of artificial intelligent systems such as large language models. And a plurality of computing chips in the server realize data interconnection through the Ethernet chip. In the conventional ethernet-based packet format for AI chip interconnection transport layer protocol, an additional TL Hdr field is added after the field of the ethernet packet header, where the TL Hdr field includes a memory operation type, a source chip physical memory address, a destination chip physical memory address, and the like. The ethernet packet header generally occupies 14 bytes, and the ethernet packet header is only used for forwarding packets in the ethernet switch chip, where no information related to the payload is carried, and for small-scale ethernet networks, for example, when only tens to hundreds of AI chips are included in the AI server, IP routing is not needed to implement data interconnection between the chips, which results in that addresses occupied by the ethernet packet header are wasted, and data transmission efficiency is low. In order to solve the above problems, when the data interconnection between the computing chips is realized through the ethernet frame, the physical memory address of the destination computing chip and the physical memory address of the source chip are respectively mapped to the destination MAC address and the source MAC address which are used for only identifying the port of the ethernet card in the original ethernet packet header, so that the field of the ethernet frame header which is originally invalid overhead is changed into the effective field