VRB-based multi-protocol intercommunication fusion method, apparatus and device, and storage medium

Abstract

The embodiment of the invention provides a VRB-based multi-protocol intercommunication fusion method, apparatus and device, and a storage medium, the method comprising: receiving a first data packet sent by an RDMA device, the first data packet comprising first data header information and first data content of an RDMA protocol; packaging an articulated naturality web protocol for the first data packet to obtain a second data packet, the second data packet including first articulated naturality web protocol information, first data header information and first data content; and sending a second data packet to a second articulated naturality web device based on the first articulated naturality web protocol information, so that the second articulated naturality web device transmits the first data content to a first large model deployed on the second articulated naturality web device according to at least part of information in the first data header information. Therefore, according to the embodiment of the invention, communication between devices with different communication protocols in a large model cluster can be realized.

Description

Technical Field

[0001] The present application relates to the field of communication technologies, and in particular, to a multi - protocol interworking and fusion method, apparatus, device, and storage medium based on VRB. Background Art

[0002] A large - model cluster refers to a system composed of multiple computing nodes. Each node has powerful computing capabilities and is interconnected through a high - speed network to jointly complete complex computing tasks. The large - model cluster is mainly used to support large - scale deep - learning model training and inference, and can be applied to fields such as artificial intelligence, big data analysis, and scientific computing.

[0003] Currently, the large - model cluster can use distributed computing technology to decompose the training task into multiple subtasks and allocate them to different computing nodes for parallel processing. Among them, communication is an essential link during distributed training in the large - model cluster. However, currently, each node in the large - model cluster only supports communication based on one communication protocol, which makes it impossible to deploy devices with different communication protocols in the large - model cluster.

[0004] It can be seen from this that in the prior art, it is impossible to achieve communication between devices with different communication protocols in the large - model cluster. Summary of the Invention

[0005] Embodiments of the present application provide a multi - protocol interworking and fusion method, apparatus, device, and storage medium based on VRB to achieve communication between devices with different communication protocols in the large - model cluster.

[0006] In a first aspect, embodiments of the present application provide a multi - protocol interworking and fusion method based on VRB, which is applied to a first Visual Networking device. The method includes:

[0007] Receiving a first data packet sent by an RDMA device, where the first data packet includes first datagram header information of the RDMA protocol and first data content;

[0008] Encapsulating a Visual Networking protocol for the first data packet to obtain a second data packet, where the second data packet includes first Visual Networking protocol information, the first datagram header information, and the first data content;

[0009] Based on the first Visual Networking protocol information, sending the second data packet to a second Visual Networking device so that the second Visual Networking device transmits the first data content to a first large - model deployed on the second Visual Networking device according to at least part of the information in the first datagram header information.

[0010] Second aspect, an embodiment of the present application provides a multi - protocol inter - communication and fusion method based on VRB, which is applied to a second Visual Networking device. The method includes:

[0011] Receiving a second data packet sent by a first Visual Networking device based on first Visual Networking protocol information, where the second data packet includes the first Visual Networking protocol information, first datagram header information of the RDMA protocol, and first data content;

[0012] Transmitting the first data content to a first large model deployed on the second Visual Networking device according to at least part of the information in the first datagram header information.

[0013] Third aspect, an embodiment of the present application provides a multi - protocol inter - communication and fusion device based on VRB, which is applied to a first Visual Networking device. The device includes:

[0014] A first receiving module, configured to receive a first data packet sent by an RDMA device, where the first data packet includes first datagram header information of the RDMA protocol and first data content;

[0015] A first processing module, configured to encapsulate the first data packet with a Visual Networking protocol to obtain a second data packet, where the second data packet includes first Visual Networking protocol information, the first datagram header information, and the first data content;

[0016] A first sending module, configured to send the second data packet to a second Visual Networking device based on the first Visual Networking protocol information, so that the second Visual Networking device transmits the first data content to a first large model deployed on the second Visual Networking device according to at least part of the information in the first datagram header information.

[0017] Fourth aspect, an embodiment of the present application provides a multi - protocol inter - communication and fusion device based on VRB, which is applied to a second Visual Networking device. The device includes:

[0018] A second receiving module, configured to receive a second data packet sent by a first Visual Networking device based on first Visual Networking protocol information, where the second data packet includes the first Visual Networking protocol information, first datagram header information of the RDMA protocol, and first data content;

[0019] A second sending module, configured to transmit the first data content to a first large model deployed on the second Visual Networking device according to at least part of the information in the first datagram header information.

[0020] Fifth aspect, an embodiment of the present application provides an electronic device, including:

[0021] One or more processors; and

[0022] One or more machine-readable media storing instructions, which, when executed by the one or more processors, cause the electronic device to execute the VRB-based multi-protocol interworking and fusion method described in the first aspect above, or execute the VRB-based multi-protocol interworking and fusion method described in the second aspect above.

[0023] In a sixth aspect, an embodiment of the present application provides a computer-readable storage medium, and the stored computer program causes a processor to execute the VRB-based multi-protocol interworking and fusion method described in the first aspect above, or execute the VRB-based multi-protocol interworking and fusion method described in the second aspect above.

[0024] In some embodiments of the present application, a first Visual Networking device can receive a first data packet sent by an RDMA device, where the first data packet includes first datagram header information of the RDMA protocol and first data content; thereby encapsulating the Visual Networking protocol for the first data packet to obtain a second data packet, where the second data packet includes first Visual Networking protocol information, first datagram header information, and first data content; and then, based on the first Visual Networking protocol information, sending the second data packet to the second Visual Networking device, so that the second Visual Networking device transmits the first data content to a first large model deployed on the second Visual Networking device according to at least part of the information in the first datagram header information.

[0025] It can be seen that in some embodiments of the present application, a Visual Networking device and an RDMA device can be deployed in a large model cluster, and in the process of transmitting the first data content of the RDMA device to the second Visual Networking device, the Visual Networking protocol information can be added to the original first data packet of the RDMA protocol through the first Visual Networking device to obtain a second data packet, and after the second Visual Networking device receives the second data packet, the first data content is transmitted to the first large model for processing based on at least part of the information in the original first datagram header information of the first data packet.

[0026] It can be known from this that in some embodiments of the present application, by encapsulating the Visual Networking protocol information for the data packet of the RDMA protocol through the first Visual Networking device, the data transmission between the RDMA device in the large model cluster and the second Visual Networking device can be realized, and the interworking and fusion of multiple protocols in the large model can be realized. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 is a flowchart of the steps of a VRB-based multi-protocol interworking and fusion method of the present application;

[0028] Figure 2 is a flowchart of the steps of another VRB-based multi-protocol interworking and fusion method of the present application;

[0029] Figure 3 It is a schematic diagram of the architecture applicable to the VRB-based multi-protocol interworking and fusion method according to an embodiment of the present application;

[0030] Figure 4 It is a schematic diagram of the transmission process of the first data content in the VRB-based multi-protocol interworking and fusion method according to an embodiment of the present application;

[0031] Figure 5 It is a schematic diagram of the transmission process of the second data content in the VRB-based multi-protocol interworking and fusion method according to an embodiment of the present application;

[0032] Figure 6 It is a schematic diagram of the transmission process of the third data content in the VRB-based multi-protocol interworking and fusion method according to an embodiment of the present application;

[0033] Figure 7 It is a schematic diagram of the specific implementation manner of the VRB-based multi-protocol interworking and fusion method according to an embodiment of the present application;

[0034] Figure 8 It is a structural block diagram of a VRB-based multi-protocol interworking and fusion device according to the present application;

[0035] Figure 9 It is a structural block diagram of another VRB-based multi-protocol interworking and fusion device according to the present application. Specific implementation manner

[0036] To make the above objects, features, and advantages of the present application more obvious and understandable, the present application will be further described in detail below with reference to the accompanying drawings and specific implementation manners.

[0037] In the prior art, each node in the large model cluster only supports communication based on one communication protocol, so that devices with different communication protocols cannot be deployed in the large model cluster.

[0038] In some embodiments of the present application, Visual Networking devices and RDMA devices can be deployed in the large model cluster. And in the process of transmitting the first data content of the RDMA device to the second Visual Networking device, the Visual Networking protocol information can be added to the basis of the first data packet of the original RDMA protocol through the first Visual Networking device to obtain the second data packet. After the second Visual Networking device receives the second data packet, the first data content is transmitted to the first large model for processing based on at least part of the information in the original first data header information of the first data packet.

[0039] The VRB-based multi-protocol interworking and fusion method according to the embodiments of the present application will be introduced in detail below.

[0040] In a first aspect, an embodiment of the present application provides a multi-protocol interoperability and integration method based on VRB; wherein, VRB is a Remote Direct Memory Access (RDMA) technology system implemented by the V2V communication protocol; that is, the VRB protocol is a remote data reading protocol provided by the Visual Networking for supporting big data transmission in the intelligent computing center.

[0041] The multi-protocol interoperability and integration method based on VRB can be applied to a first Visual Networking device, such as Figure 1 As shown, the method may include the following steps 101 to 103:

[0042] Step 101: Receive a first data packet sent by an RDMA device.

[0043] Wherein, the first data packet includes first datagram header information of the RDMA protocol and first data content. Optionally, the first datagram header information may include an Internet Protocol (IP) header and a User Datagram Protocol (UDP) header.

[0044] The IP header includes a source IP address and a destination IP address; for example, when the first data packet needs to be sent from an RDMA device to a first Visual Networking device, and then the first Visual Networking device sends the first data content in the first data packet to a second Visual Networking device, the source IP address in the IP header of the first data packet is the IP address of the RDMA device, and the destination IP address is the IP address of the second Visual Networking device.

[0045] The UDP header is a part of the UDP protocol for transmitting data packets, including some key information for ensuring the correct transmission and processing of data packets.

[0046] Optionally, the first data packet further includes a source physical address and a destination physical address; for example, when the first data packet needs to be sent from an RDMA device to a first Visual Networking device, and then the first Visual Networking device further transmits the first data content in the first data packet to a second Visual Networking device, the source physical address in the first data packet is the physical address of the RDMA device, and the destination physical address is the physical address of the first Visual Networking device. Wherein, the physical address may be a Media Access Control (MAC) address.

[0047] In addition, the first data content can be understood as: the data content that needs to be transmitted to the second Visual Networking device for processing by the first large model deployed on the second Visual Networking device; that is, it can be understood that the first data content includes the input content of the first large model. It can be seen that in some embodiments of the present application, when the first data content of the RDMA device needs to be processed by the first large model deployed on the second Visual Networking device, the RDMA device can first transmit the first data content to the first Visual Networking device, and then the first Visual Networking device transmits it to the second Visual Networking device.

[0048] Step 102: Encapsulate the Visual Networking protocol for the first data packet to obtain a second data packet.

[0049] Wherein, the second data packet includes first Visual Networking protocol information, the first data header information, and the first data content.

[0050] As can be seen from step 102, after receiving the first data packet of the RDMA protocol sent by the RDMA device, the first Visual Networking device can further encapsulate the Visual Networking protocol on the basis of the first data packet, so as to obtain a second data packet with the first Visual Networking protocol information added. In this way, the first Visual Networking device can send the second data packet to the second Visual Networking device based on the first Visual Networking protocol information.

[0051] Optionally, the first Visual Networking protocol information includes a source Visual Networking number and a destination Visual Networking number; for example, when the second data packet needs to be sent from the first Visual Networking device to the second Visual Networking device, the source Visual Networking number in the first Visual Networking protocol information is the Visual Networking number of the first Visual Networking device, and the destination Visual Networking number is the Visual Networking number of the second Visual Networking device.

[0052] Optionally, the second data packet further includes a source physical address and a destination physical address; for example, when the second data packet needs to be sent from the first Visual Networking device to the second Visual Networking device, the source physical address in the second data packet is the physical address of the first Visual Networking device, and the destination physical address is the physical address of the second Visual Networking device.

[0053] Step 103: Based on the first Visual Networking protocol information, send the second data packet to the second Visual Networking device, so that the second Visual Networking device transmits the first data content to the first large model deployed on the second Visual Networking device according to at least part of the information in the first data header information.

[0054] As can be seen from step 102 above, after the first Visual Networking device receives the first data packet of the RDMA protocol sent by the RDMA device, the first Visual Networking protocol information is further added to the basis of the first data packet, that is, the original first data header information of the RDMA protocol in the first data packet is retained. In this way, the original first data header information exists in the second data packet of the Visual Networking protocol. As a result, after the second Visual Networking device receives the second data packet, it does not need to convert the relevant information of the RDMA protocol based on the first Visual Networking protocol information, and can directly transmit the first data content to the first large model deployed on the second Visual Networking device according to at least some of the information in the first data header information.

[0055] Exemplarily, when the first data header information includes an IP header and a UDP header, the IP header includes a source IP address (i.e., the IP address of the RDMA device) and a destination IP address (i.e., the IP address of the second Visual Networking device), and the UDP header includes check information. After the second Visual Networking device receives the second data packet, it does not need to convert the destination Visual Networking number (i.e., the Visual Networking number of the second Visual Networking device) in the first Visual Networking protocol information into an IP address, and can directly use the destination IP address, source IP address in the first data header information, and the check information in the UDP header to transmit the first data content to the first large model deployed on the second Visual Networking device.

[0056] Optionally, the first Visual Networking device can be connected to multiple Visual Networking devices. In this case, the first corresponding relationship between the physical addresses and ports of the respective Visual Networking devices connected to the first Visual Networking device can be recorded. In this way, when the second data packet includes the physical address of the second Visual Networking device, the first Visual Networking device can determine the corresponding port according to the physical address of the second Visual Networking device in the second data packet and the first corresponding relationship, and then send the second data packet to the second Visual Networking device through this port.

[0057] As can be seen from steps 101 to 103 above, in some embodiments of the present application, the first Visual Networking device can receive the first data packet sent by the RDMA device, where the first data packet includes the first data header information and the first data content of the RDMA protocol; thus encapsulating the Visual Networking protocol for the first data packet to obtain a second data packet, where the second data packet includes the first Visual Networking protocol information, the first data header information, and the first data content; and then sending the second data packet to the second Visual Networking device based on the first Visual Networking protocol information, so that the second Visual Networking device transmits the first data content to the first large model deployed on the second Visual Networking device according to at least some of the information in the first data header information.

[0058] It can be seen that in some embodiments of the present application, a Visual Networking device and an RDMA device can be deployed in a large model cluster. And during the process of transmitting the first data content of the RDMA device to the second Visual Networking device, the first Visual Networking device can add Visual Networking protocol information to the basis of the first data packet of the original RDMA protocol to obtain a second data packet. After the second Visual Networking device receives the second data packet, the first data content is transmitted to the first large model for processing based on at least part of the information in the original first data header information of the first data packet.

[0059] It can be known from this that in some embodiments of the present application, by encapsulating Visual Networking protocol information for the data packet of the RDMA protocol through the first Visual Networking device, data transmission between the RDMA device and the second Visual Networking device in the large model cluster can be realized, and the intercommunication and integration of multiple protocols in the large model can be achieved.

[0060] In some embodiments of the present application, the IP address of the second Visual Networking device is included in the first data header information; the method further includes the following steps A-1 to A-2:

[0061] Step A-1: Obtain the physical address of the second Visual Networking device according to the IP address of the second Visual Networking device;

[0062] Step A-2: Encapsulate the physical address of the second Visual Networking device into the second data packet;

[0063] In the above step 103, sending the second data packet to the second Visual Networking device based on the first Visual Networking protocol information includes the following steps B-1:

[0064] Step B-1: Send the second data packet to the second Visual Networking device based on the first Visual Networking protocol information and the physical address of the second Visual Networking device.

[0065] It can be known from this that in the case where the IP address of the second Visual Networking device is included in the first data header information of the first data packet, the first Visual Networking device can obtain the physical address of the second Visual Networking device according to this IP address, so as to also encapsulate the physical address of the second Visual Networking device into the second data packet. In this way, the first Visual Networking device can send the second data packet to the second Visual Networking device based on the first Visual Networking protocol information and according to the physical address of the second Visual Networking device.

[0066] Among them, the first Visual Networking device can pre-record the second corresponding relationship between the physical address and the IP address of each Visual Networking device accessing it. In this way, the first Visual Networking device can obtain the physical address of the second Visual Networking device based on the IP address of the second Visual Networking device and the second corresponding relationship.

[0067] In some embodiments of the present application, after the first data content is transmitted to the first large model, the method further includes the following steps C-1 to C-3:

[0068] Step C-1: Receive a third data packet sent by the second Visual Networking device based on the second Visual Networking protocol information, where the third data packet includes the second Visual Networking protocol information, the second data header information of the RDMA protocol, and the second data content, and the second data content includes: the content output by the first large model after the first data content is input into the first large model;

[0069] Step C-2: Remove the second Visual Networking protocol information from the third data packet to obtain a fourth data packet, where the fourth data packet includes the second data header information and the second data content;

[0070] Step C-3: Based on the second data header information, send the fourth data packet to the RDMA device.

[0071] Among them, after the first data content is transmitted to the first large model deployed on the second Visual Networking device, the second data content can be output, and this second data content can then be returned to the RDMA device.

[0072] In some embodiments of the present application, after the second Visual Networking device obtains the second data content, it can encapsulate the RDMA protocol and the Visual Networking protocol for the second data content (that is, it can generate the second data header information of the RDMA protocol for the second data content and encapsulate the second Visual Networking protocol information), thereby obtaining the above-mentioned third data packet, and then based on the second Visual Networking protocol information, send the third data packet to the first Visual Networking device, so that the first Visual Networking device can then send the second data content therein to the RDMA device.

[0073] It can be seen from the above steps C-1 to C-3 that after the first Visual Networking device receives the third data packet, it can remove the second Visual Networking protocol information from the third data packet, thereby obtaining the above-mentioned fourth data packet, and then based on the second data header information of the RDMA protocol in the fourth data packet, send the fourth data packet to the RDMA device. In this way, the second data content output by the first large model is returned to the RDMA device.

[0074] Optionally, the second Visual Networking protocol information includes a source Visual Networking number and a destination Visual Networking number; for example, when the third data packet needs to be sent from the second Visual Networking device to the first Visual Networking device, the source Visual Networking number in the second Visual Networking protocol information is the Visual Networking number of the second Visual Networking device, and the destination Visual Networking number is the Visual Networking number of the first Visual Networking device.

[0075] Optionally, the second data header information may include an IP header and a UDP header. The IP header includes a source IP address and a destination IP address; for example, when the third data packet needs to be sent from the second Visual Networking device to the first Visual Networking device, and then the first Visual Networking device sends the second data content in the third data packet to the RDMA device, the source IP address in the IP header of the second data header information is the IP address of the second Visual Networking device, and the destination IP address is the IP address of the RDMA device. It should be noted that in this case, the IP address of the RDMA device in the IP header of the second data header information can be extracted and recorded by the second Visual Networking device from the first data header information in the second data packet after receiving the second data packet.

[0076] Optionally, the third data packet further includes a source physical address and a destination physical address; for example, when the third data packet needs to be sent from the second Visual Networking device to the first Visual Networking device, and then the first Visual Networking device sends the second data content in the third data packet to the RDMA device, the source physical address in the third data packet is the physical address of the second Visual Networking device, and the destination physical address is the physical address of the first Visual Networking device.

[0077] Optionally, the fourth data packet further includes a source physical address and a destination physical address; for example, when the fourth data packet needs to be sent from the first Visual Networking device to the RDMA device, the source physical address in the fourth data packet is the physical address of the first Visual Networking device, and the destination physical address is the physical address of the RDMA device.

[0078] Optionally, the first Visual Networking device may be connected to multiple RDMA devices. In this case, the first Visual Networking device may record the third correspondence between the physical addresses and ports of each RDMA device. In this way, the first Visual Networking device can determine the corresponding port according to the physical address of the RDMA device in the fourth data packet and the third correspondence, and then send the fourth data packet to the RDMA device through this port.

[0079] In some embodiments of the present application, the IP address of the RDMA device is included in the second data header information; the method further includes the following steps D-1 to D-2:

[0080] Step D-1: Obtain the physical address of the RDMA device according to the IP address of the RDMA device;

[0081] Step D-2: Enclose the physical address of the RDMA device into the fourth data packet;

[0082] In the above step C-4, based on the second datagram header information, sending the fourth data packet to the RDMA device includes the following steps C-3.1:

[0083] Step C-3.1: Based on the physical address of the RDMA device and the second datagram header information, send the fourth data packet to the RDMA device.

[0084] It can be seen that in the case where the IP address of the RDMA device is included in the second datagram header information of the third data packet, the first Visual Networking device can obtain the physical address of the RDMA device according to the IP address, so that the physical address of the RDMA device is also encapsulated into the fourth data packet. In this way, the first Visual Networking device can send the fourth data packet to the RDMA device based on the second datagram header information and the physical address of the RDMA device.

[0085] Among them, the first Visual Networking device can pre-record the fourth correspondence between the physical addresses and IP addresses of each RDMA device accessing it. In this way, the first Visual Networking device can obtain the physical address of the RDMA device based on the above IP address of the RDMA device and the fourth correspondence.

[0086] In some embodiments of the present application, the method further includes the following steps D-1 to D-4:

[0087] Step D-1: Receive the fifth data packet sent by the second Visual Networking device based on the third Visual Networking protocol information, where the fifth data packet includes the third Visual Networking protocol information, the physical address of the RDMA device, and third data content;

[0088] Step D-2: Determine the IP address of the RDMA device according to the physical address of the RDMA device;

[0089] Step D-3: Remove the third Visual Networking protocol information in the fifth data packet, and process the fifth data packet according to the IP address of the RDMA device to obtain a sixth data packet, where the sixth data packet includes the physical address of the RDMA device, the third datagram header information of the RDMA protocol, and the third data content, and the third datagram header information includes the IP address of the RDMA device;

[0090] Step D-4: Based on the physical address of the RDMA device and the third datagram header information, send the sixth data packet to the RDMA device, so that the RDMA device transmits the third data content to the second large model deployed on the RDMA device according to at least part of the information in the third datagram header information.

[0091] It should be noted that the above third data content can be understood as the data content actively sent by the second Visual Networking device to the RDMA device, that is, it does not belong to the output content of the first large model; for example, the third data content can include the content that needs to be sent to the RDMA device for processing by the second large model deployed on the RDMA device, that is, the third data content can include the input content of the second large model.

[0092] In some embodiments of the present application, the second Visual Networking device can encapsulate the physical address of the RDMA device and the third Visual Networking protocol information for the third data content, so as to obtain the above fifth data packet, and then send the fifth data packet to the first Visual Networking device based on the third Visual Networking protocol information, so that the first Visual Networking device can then send the third data content therein to the RDMA device.

[0093] As can be seen from the above steps D-1 to D-4, after receiving the fifth data packet, the first Visual Networking device can determine the IP address of the RDMA device according to the physical address of the RDMA device, remove the third Visual Networking protocol information in the fifth data packet, and obtain the above third data header information according to the IP address of the RDMA device. Then, the physical address of the RDMA device, the third data header information, and the third data content are encapsulated to obtain the above sixth data packet, so that the first Visual Networking device sends the sixth data packet to the RDMA device according to the physical address of the RDMA device and the third data header information, so that the RDMA device can transmit the third data content to the second large model deployed on the RDMA device according to at least part of the information in the third data header information.

[0094] Optionally, the third data header information may include an IP header and a UDP header. The IP header includes a source IP address and a destination IP address; for example, when the fifth data packet needs to be sent from the second Visual Networking device to the first Visual Networking device, and then the first Visual Networking device sends the third data content in the fifth data packet to the RDMA device, the source IP address in the IP header of the third data header information is the IP address of the second Visual Networking device, and the destination IP address is the IP address of the RDMA device.

[0095] Exemplarily, when the third data header information includes an IP header and a UDP header, the IP header includes a source IP address (i.e., the IP address of the second Visual Networking device) and a destination IP address (i.e., the IP address of the RDMA device), and the UDP header includes check information. After receiving the sixth data packet, the RDMA device can use the destination IP address in the third data packet header and the check information in the UDP header to transmit the third data content to the second large model deployed on the first Visual Networking device.

[0096] It should be noted that the process of the fourth data content being output and returned to the second Visual Networking device after the third data content is input into the second large model is the same as the process of the first data content being sent from the RDMA device to the second Visual Networking device described above, and will not be elaborated here.

[0097] Optionally, the above Visual Networking protocol may be a VRB protocol, and the above RDMA protocol may be a RoCE v2 protocol.

[0098] Among them, the Remote Direct Memory Access over Converged Ethernet version 2 (RoCE v2) protocol is designed to achieve high-performance and low-latency data transmission through an Ethernet network. RoCE v2 combines Ethernet and RDMA technologies, significantly enhancing performance and flexibility, and is particularly suitable for application scenarios with extremely high requirements for data interaction speed and efficiency.

[0099] In some embodiments of the present application, the first Visual Networking device and the second Visual Networking device are VRB devices.

[0100] In a second aspect, an embodiment of the present application provides a multi-protocol interoperability and fusion method based on VRB; applied to the second Visual Networking device, as Figure 2 shown, the method may include the following steps 201 to 202:

[0101] Step 201: Receive a second data packet sent by the first Visual Networking device based on the first Visual Networking protocol information.

[0102] Among them, the second data packet includes the first Visual Networking protocol information, the first data header information of the RDMA protocol, and the first data content.

[0103] In addition, after the first Visual Networking device receives the first data packet (including the first data header information and the first data content of the RDMA protocol) sent by the RDMA device, it encapsulates the Visual Networking protocol for the first data packet to obtain a second data packet, and thus sends the second data packet to the second Visual Networking device based on the first Visual Networking protocol information in the second data packet.

[0104] In addition, the first data content can be understood as: the data content that needs to be transmitted to the second Visual Networking device for processing by the first large model deployed on the second Visual Networking device; that is, it can be understood that the first data content includes the input content of the first large model. It can be seen that in some embodiments of the present application, when the first data content of the RDMA device needs to be processed by the first large model deployed on the second Visual Networking device, the RDMA device can first transmit the first data content to the first Visual Networking device, and then the first Visual Networking device transmits it to the second Visual Networking device.

[0105] Optionally, the first Visual Networking protocol information includes a source Visual Networking number and a destination Visual Networking number; for example, when the second data packet needs to be sent from the first Visual Networking device to the second Visual Networking device, the source Visual Networking number in the first Visual Networking protocol information is the Visual Networking number of the first Visual Networking device, and the destination Visual Networking number is the Visual Networking number of the second Visual Networking device.

[0106] Optionally, the second data packet further includes a source physical address and a destination physical address; for example, when the second data packet needs to be sent from the first Visual Networking device to the second Visual Networking device, the source physical address in the second data packet is the physical address of the first Visual Networking device, and the destination physical address is the physical address of the second Visual Networking device.

[0107] Step 202: Transmit the first data content to the first large model deployed on the second Visual Networking device according to at least part of the information in the first data header information.

[0108] As can be seen from the above, after the first Visual Networking device receives the first data packet of the RDMA protocol sent by the RDMA device, the first Visual Networking protocol information is further added on the basis of the first data packet, that is, the original first data header information of the RDMA protocol in the first data packet is retained. In this way, the original first data header information exists in the second data packet of the Visual Networking protocol, so that after the second Visual Networking device receives the second data packet, it does not need to convert the relevant information of the RDMA protocol based on the first Visual Networking protocol information, and can directly transmit the first data content to the first large model deployed on the second Visual Networking device according to at least part of the information in the first data header information.

[0109] Exemplarily, when the first datagram header information includes an IP header and a UDP header, the IP header includes a source IP address (i.e., the IP address of the RDMA device) and a destination IP address (i.e., the IP address of the second Visual Networking device), and the UDP header includes check information. After the second Visual Networking device receives the second data packet, it does not need to convert the destination Visual Networking number (i.e., the Visual Networking number of the second Visual Networking device) in the first Visual Networking protocol information into an IP address. Instead, it can directly use the destination IP address in the first datagram header information, the source IP address, and the check information in the UDP header to transmit the first data content to the first large model deployed on the second Visual Networking device.

[0110] As can be seen from the above steps 201 to 202, in some embodiments of the present application, the second Visual Networking device can receive the second data packet sent by the first Visual Networking device based on the first Visual Networking protocol information. The second data packet includes the first Visual Networking protocol information, the first datagram header information of the RDMA protocol, and the first data content. Then, according to at least some of the information in the first datagram header information, the first data content is transmitted to the first large model deployed on the second Visual Networking device.

[0111] It can be seen that in some embodiments of the present application, Visual Networking devices and RDMA devices can be deployed in the large model cluster. During the process of transmitting the first data content of the RDMA device to the second Visual Networking device, the first Visual Networking device can add Visual Networking protocol information to the original first data packet of the RDMA protocol to obtain the second data packet. After the second Visual Networking device receives the second data packet, based on at least some of the information in the original first datagram header information of the first data packet, the first data content is transmitted to the first large model for processing.

[0112] From this, it can be known that in some embodiments of the present application, by encapsulating Visual Networking protocol information for the data packet of the RDMA protocol through the first Visual Networking device, data transmission between the RDMA device in the large model cluster and the second Visual Networking device can be achieved, realizing the intercommunication and integration of multiple protocols in the large model.

[0113] In some embodiments of the present application, the first datagram header information includes the IP address of the second Visual Networking device; the second data packet further includes: the physical address of the second Visual Networking device determined according to the IP address of the second Visual Networking device.

[0114] Wherein, when the IP address of the second Visual Networking device is included in the first datagram header information of the first data packet, the first Visual Networking device may obtain the physical address of the second Visual Networking device according to the IP address, and then encapsulate the physical address of the second Visual Networking device into the second data packet. In this way, the first Visual Networking device may send the second data packet to the second Visual Networking device based on the first Visual Networking protocol information and according to the physical address of the second Visual Networking device.

[0115] In some embodiments of the present application, the method further includes the following steps E-1 to E-2:

[0116] Step E-1: After obtaining the second data content, encapsulate the Visual Networking protocol and the RDMA protocol for the second data content to obtain a third data packet, wherein the second data content includes the content output by the first large model after the first data content is input into the first large model, and the third data packet includes second Visual Networking protocol information, second datagram header information of the RDMA protocol, and the second data content;

[0117] Step E-2: Send the third data packet to the first Visual Networking device based on the second Visual Networking protocol information, so that the first Visual Networking device removes the second Visual Networking protocol information in the third data packet to obtain a fourth data packet, and then sends the fourth data packet to the RDMA device based on the second data packet header, wherein the fourth data packet includes the second datagram header information and the second data content.

[0118] Wherein, after the first data content is transmitted to the first large model deployed on the second Visual Networking device, the second data content may be output, and the second data content may then be returned to the RDMA device.

[0119] It can be seen from the above steps E-1 to E-2 that after the second Visual Networking device obtains the second data content, it can encapsulate the RDMA protocol and the Visual Networking protocol for the second data content (that is, it can generate the second datagram header information of the RDMA protocol for the second data content and encapsulate the second Visual Networking protocol information), thereby obtaining the above-mentioned third data packet, and then sending the third data packet to the first Visual Networking device based on the second Visual Networking protocol information, so that the first Visual Networking device then sends the second data content therein to the RDMA device.

[0120] In addition, in some embodiments of the present application, after receiving the third data packet, the first Visual Networking device may remove the second Visual Networking protocol information in the third data packet, thereby obtaining the above-mentioned fourth data packet, and then send the fourth data packet to the RDMA device based on the second datagram header information of the RDMA protocol in the fourth data packet. In this way, the second data content output by the first large model is returned to the RDMA device.

[0121] In some embodiments of the present application, the first data header information further includes the IP address of the RDMA device; the method further includes the following step F-1:

[0122] Step F-1: Extract the IP address of the RDMA device from the first data header information and record it.

[0123] In the above step E-1, encapsulating the RDMA protocol for the second data content includes the following step E-1.1:

[0124] Step E-1.1: According to the IP address of the RDMA device, encapsulate the RDMA protocol for the second data content to obtain the second data header information including the IP address of the RDMA device.

[0125] It can be seen that after the second Visual Networking device receives the second data packet, it can extract the IP address of the RDMA device from the first data header information of the second data packet and record it. Thus, after the first data content is input into the first large model to obtain the second data content, the RDMA protocol can be encapsulated for the second data content according to the IP address of the RDMA device to obtain the second data header information including the IP address of the RDMA device.

[0126] In some embodiments of the present application, the method further includes the following step G-1:

[0127] Step G-1: Establish a connection channel with the first Visual Networking device.

[0128] In the above step F-1, extracting the IP address of the RDMA device from the first data header information includes the following step F-1.1:

[0129] Step F-1.1: In the connection channel for receiving the second data packet, when it is the connection channel between the second Visual Networking device and the first Visual Networking device, extract the IP address of the RDMA device from the first data header information.

[0130] It can be seen therefrom that a connection channel can be established between the first Visual Networking device and the second Visual Networking device, so as to realize data transmission between the first Visual Networking device and the second Visual Networking device through this connection channel. Moreover, the second Visual Networking device can determine from which device the received data channel comes through the connection channel for receiving data packets; wherein, if it is recognized that a data packet sent by the first Visual Networking device is received, the data packet may include the data content forwarded by the first Visual Networking device for the RDMA device, that is, the IP address of the RDMA device may exist in the data packet, so that the IP address of the RDMA device therein can be extracted, facilitating the second Visual Networking device to return data to the RDMA device.

[0131] Optionally, one Visual Networking number can correspond to multiple sub-device numbers, that is, one Visual Networking device can have multiple sub-device numbers. When establishing a connection channel between two Visual Networking devices, one sub-device number is required respectively. Therefore, when establishing a connection channel, the first Visual Networking device and the second Visual Networking device can negotiate which sub-device number to use to avoid the same sub-device number being occupied by multiple channels.

[0132] Optionally, after the second Visual Networking device establishes a connection channel with the first Visual Networking device, it can record the corresponding relationship between the sub-device number corresponding to this connection channel and the device identification information of the first Visual Networking device. In this way, the sub-device number can be used to identify the connection channel, so that based on this corresponding relationship, it can be recognized whether the device corresponding to the connection channel for receiving data packets is the first Visual Networking device.

[0133] In some embodiments of the present application, the method further includes the following step H-1:

[0134] Step H-1: Send a fifth data packet to the first Visual Networking device based on the third Visual Networking protocol information, so that the first Visual Networking device determines the IP address of the RDMA device according to the physical address of the RDMA device, removes the third Visual Networking protocol information in the fifth data packet, processes the fifth data packet according to the IP address of the RDMA device to obtain a sixth data packet, and sends the sixth data packet to the RDMA device based on the physical address of the RDMA device and the third data header information of the RDMA protocol;

[0135] Wherein, the fifth data packet includes the third Visual Networking protocol information, the physical address of the RDMA device, and the third data content;

[0136] The sixth data packet includes the physical address of the RDMA device, the third data header information, and the third data content, and the third data header information includes the IP address of the RDMA device.

[0137] It should be noted that the above third data content can be understood as the data content actively sent by the second Visual Networking device to the RDMA device, that is, it does not belong to the output content of the first large model; for example, the third data content can include the content that needs to be sent to the RDMA device for processing by the second large model deployed on the RDMA device, that is, the third data content can include the input content of the second large model.

[0138] As can be seen from the above step H-1, the second Visual Networking device can encapsulate the physical address of the RDMA device and the third Visual Networking protocol information for the third data content, so as to obtain the above fifth data packet, and then send the fifth data packet to the first Visual Networking device based on the third Visual Networking protocol information, so that the first Visual Networking device can send the third data content therein to the RDMA device.

[0139] In some embodiments of the present application, after receiving the fifth data packet, the first Visual Networking device can determine the IP address of the RDMA device according to the physical address of the RDMA device, remove the third Visual Networking protocol information in the fifth data packet, and obtain the above third data header information according to the IP address of the RDMA device, and then encapsulate the physical address of the RDMA device, the third data header information, and the third data content to obtain the above sixth data packet, so that the first Visual Networking device sends the sixth data packet to the RDMA device according to the physical address of the RDMA device and the third data header information, so that the RDMA device transmits the third data content to the second large model deployed on the RDMA device according to at least part of the information in the third data header information.

[0140] Optionally, the third data header information may include an IP header and a UDP header. The IP header includes a source IP address and a destination IP address; for example, when the fifth data packet needs to be sent from the second Visual Networking device to the first Visual Networking device, and then the first Visual Networking device sends the third data content in the fifth data packet to the RDMA device, the source IP address in the IP header of the third data header information is the IP address of the second Visual Networking device, and the destination IP address is the IP address of the RDMA device.

[0141] Exemplarily, when the third data header information includes an IP header and a UDP header, the IP header includes a source IP address (i.e., the IP address of the second Visual Networking device) and a destination IP address (i.e., the IP address of the RDMA device), and the UDP header includes check information, then after receiving the sixth data packet, the RDMA device can use the destination IP address in the third data packet header and the check information in the UDP header to transmit the third data content to the second large model deployed on the first Visual Networking device.

[0142] It should be noted that the process of inputting the third data content into the second large model and returning the output fourth data content to the second Visual Networking device is the same as the process of the first data content being sent from the RDMA device to the second Visual Networking device described above, and will not be elaborated here.

[0143] In summary, the specific implementation of the VRB-based multi-protocol interworking and fusion method in the embodiments of this application can be described as follows:

[0144] The VRB-based multi-protocol interworking and fusion method in this implementation can be applied to an architecture as Figure 3 shown. Among them, based on the architecture as Figure 3 shown, the specific implementation process of this VRB-based multi-protocol interworking and fusion method is described as follows:

[0145] (1) The Visual Networking protocol stack of the VRB device and the Visual Networking proxy of the X-VRB device need to access the Visual Networking. Among them, according to the configured MAC address, access to the Visual Networking can be applied for to obtain a Visual Networking number. Taking the Visual Networking proxy of the X-VRB device as an example below, the process of accessing the Visual Networking is introduced:

[0146] The Visual Networking server can send a broadcast data packet, and at least one pre-determined MAC can be carried in this broadcast data packet. For example, the carried MAC is MAC1. Among them, if the Visual Networking proxy needs to access the network through MAC1, it can encapsulate a Visual Networking network access application packet to the Visual Networking server according to MAC1. The Visual Networking server judges the legality of the MAC1 of the Visual Networking proxy (that is, judges whether the MAC in the network access application packet is the same as the locally stored MAC1). If it is legal (that is, the MAC in the network access application packet is the same as the locally stored MAC1), the Visual Networking number is returned.

[0147] (2) The Visual Networking protocol stack of the VRB device and the Visual Networking proxy of the X-VRB device negotiate a connection channel. After the connection channel between the two is established, the X-VRB device informs the VRB device of the device identifier of the X-VRB device, and then the VRB device saves the information of the connection channel and the corresponding device identifier of the X-VRB device to the Visual Networking protocol stack.

[0148] It should be noted that: both the VRB device and the X-VRB device belong to Visual Networking devices. Each Visual Networking device has a Visual Networking number, and one Visual Networking number corresponds to multiple sub-device numbers, that is, one Visual Networking device will have multiple sub-device numbers. Among them, when establishing a channel between two Visual Networking devices, a sub-device number needs to be used respectively. Therefore, when establishing a channel, it is necessary to negotiate which sub-device number to use to avoid the same sub-device number being occupied by multiple channels.

[0149] (3) When the first data content of the RDMA device needs to be transmitted to the VRB device and processed by the first large model deployed on the VRB device, the transmission process of the first data content is as follows Figure 4 as shown:

[0150] Based on the first data content, the RDMA device encapsulates the first data packet. The first data packet includes the source MAC address (i.e., the MAC address of the RDMA device) and the destination MAC address (i.e., the MAC address of the X-VRB device), an IP header, a UDP header, and the first data content. The IP header includes the source IP address (i.e., the IP address of the RDMA device) and the destination IP address (i.e., the IP address of the VRB device);

[0151] The RDMA device sends the first data packet to the network card of the X-VRB device through its network card, so that the network card of the X-VRB device transmits it to the Visual Networking Agent of the X-VRB device, and then the Visual Networking Agent transmits it to the protocol conversion unit. The protocol conversion unit encapsulates the Visual Networking protocol based on the first data packet and updates the source MAC address and the destination MAC address, thereby obtaining the second data packet. The second data packet includes: the first Visual Networking protocol information, the source MAC address (i.e., the MAC address of the X-VRB device) and the destination MAC address (i.e., the MAC address of the VRB device), an IP header, a UDP header, and the first data content. The first Visual Networking protocol information includes the source Visual Networking number (i.e., the Visual Networking number of the X-VRB device) and the destination Visual Networking number (i.e., the Visual Networking number of the VRB device). The IP header includes the source IP address (i.e., the IP address of the RDMA device) and the destination IP address (i.e., the IP address of the VRB device);

[0152] After obtaining the above second data packet, the protocol conversion unit transmits the second data packet to the Visual Networking Agent. The Visual Networking Agent transmits the second data packet to the Visual Networking protocol stack of the VRB device through the connection channel between its Visual Networking protocol stack and the VRB device (i.e., the channel formed by the Visual Networking Agent - the network card of the X-VRB device - the network card of the VRB device - the Visual Networking protocol stack);

[0153] After receiving the second data packet, the Visual Networking protocol stack extracts the source IP address (i.e., the IP address of the RDMA device) from the second data packet and saves the IP address to the queue; then, it transmits the second data packet to the kernel-mode driver of the VRB device, so that the kernel-mode driver transmits the first data content to the first large model deployed on the VRB device based on the destination IP address (i.e., the IP address of the VRB device), the source IP address (i.e., the IP address of the RDMA device), and the check information in the UDP header;

[0154] It should be noted that the kernel layer (i.e., the kernel-mode driver) needs to use the IP address of the VRB device and the IP address of the RDMA device for virtual memory and physical memory conversion; in this embodiment, the IP header memory included in the second data packet has the destination IP address (i.e., the IP address of the VRB device) and the source IP address (i.e., the IP address of the RDMA device). Therefore, after receiving the second data packet, the kernel-mode driver can directly extract the IP address of the VRB device and the IP address of the RDMA device from the second data packet for use.

[0155] In addition, after receiving the second data packet, the kernel layer needs to perform verification on it (such as verifying its integrity), and then it can perform verification according to the verification information in the UDP header of the second data packet.

[0156] (4) After the first data content is input into the first large model deployed on the VRB device and the second data content is output, when the second data content needs to be returned to the RDMA device, the transmission process of the second data content is as Figure 5 shown:

[0157] After the first large model outputs the second data content, it transmits the second data content to the kernel-mode driver of the VRB device. The kernel-mode driver generates the corresponding UDP header based on the second data content and transmits the second data content and its corresponding UDP header to the Visual Networking protocol stack;

[0158] The Visual Networking protocol stack extracts the IP address of the RDMA device from the queue and generates the IP header corresponding to the second data content based on this IP address; and generates the second Visual Networking protocol information corresponding to the second data content; based on the IP header, UDP header, and second Visual Networking protocol information corresponding to the second data content, it generates a third data packet. Among them, the third data packet includes: the second Visual Networking protocol information, the source MAC address (i.e., the MAC address of the VRB device) and the destination MAC address (i.e., the MAC address of the X-VRB device), the IP header, the UDP header, and the second data content. Among them, the second Visual Networking protocol information includes the source Visual Networking number (i.e., the Visual Networking number of the VRB device) and the destination Visual Networking number (i.e., the Visual Networking number of the X-VRB device), and the IP header includes the source IP address (i.e., the IP address of the VRB device) and the destination IP address (i.e., the IP address of the RDMA device);

[0159] After obtaining the above third data packet, the Visual Networking protocol stack transmits the third data packet to the Visual Networking proxy of the X-VRB device through the connection channel between the Visual Networking protocol stack and the Visual Networking proxy of the X-VRB device;

[0160] After receiving the third data packet, the Visual Networking proxy transmits the third data packet to the protocol conversion unit;

[0161] The protocol conversion unit obtains the MAC address corresponding to the IP address of the RDMA device in the third data packet according to the recorded correspondence between the MAC address and the IP address of the RDMA device, and uses it as the MAC address of the RDMA device; and processes the third data packet based on the MAC address of the RDMA device to obtain a fourth data packet, where the fourth data packet includes: a source MAC address (i.e., the MAC address of the VRB device) and a destination MAC address (i.e., the MAC address of the RDMA device), an IP header, a UDP header, and second data content, where the IP header includes a source IP address (i.e., the IP address of the VRB device) and a destination IP address (i.e., the IP address of the RDMA device);

[0162] After obtaining the fourth data packet, the protocol conversion unit returns it to the Visual Networking proxy, and then the Visual Networking proxy transmits it to the network card of the X-VRB device, and then the network card of the X-VRB device transmits it to the network card of the RDMA device.

[0163] (5) When the third data content of the VRB device needs to be transmitted to the RDMA device and processed by the second large model deployed on the RDMA device, the transmission process of the third data content is as follows Figure 6 shown:

[0164] The Visual Networking protocol stack encapsulates a fifth data packet based on the third data content, where the fifth data packet includes: third Visual Networking protocol information, a source MAC address (i.e., the MAC address of the VRB device) and a destination MAC address (i.e., the MAC address of the X-VRB device), an IP header, a UDP header, the MAC address of the RDMA device, and the third data content, where the third Visual Networking protocol information includes a source Visual Networking number (i.e., the Visual Networking number of the VRB device) and a destination Visual Networking number (i.e., the Visual Networking number of the X-VRB device), and the IP header includes a source IP address (i.e., the IP address of the VRB device) and a destination IP address (i.e., the IP address of the RDMA device);

[0165] After obtaining the above fifth data packet, the fifth data packet is transmitted to the Visual Networking proxy of the X-VRB device through the connection channel between the Visual Networking protocol stack and the Visual Networking proxy of the X-VRB device;

[0166] After receiving the fifth data packet, the Visual Networking proxy transmits the third data packet to the protocol conversion unit;

[0167] The protocol conversion unit obtains the IP address corresponding to the MAC address of the RDMA device in the third data packet according to the recorded correspondence between the MAC address and the IP address of the RDMA device, and uses it as the IP address of the RDMA device; and generates an IP header based on the IP address of the RDMA device; generates a UDP header, thereby obtaining a sixth data packet, where the sixth data packet includes: a source MAC address (i.e., the MAC address of the VRB device) and a destination MAC address (i.e., the MAC address of the RDMA device), an IP header, a UDP header, and the third data content, where the IP header includes a source IP address (i.e., the IP address of the VRB device) and a destination IP address (i.e., the IP address of the RDMA device);

[0168] After obtaining the sixth data packet, the protocol conversion unit returns it to the Visual Networking Agent, which is then transmitted by the Visual Networking Agent to the network card of the X-VRB device, and then transmitted by the network card of the X-VRB device to the network card of the RDMA device;

[0169] After receiving the sixth data packet, the network card of the RDMA device transmits the third data content in the sixth data packet to the second large model deployed on the RDMA device through the TCP / IP protocol stack, the kernel-mode driver, the user-mode driver, and the RDMA Application Programming Interface (API) in sequence.

[0170] In summary, as described in (1) to (5) above, the multi-protocol interworking and fusion method based on VRB in the embodiments of the present application, as Figure 7 shown, mainly includes the following three aspects of management:

[0171] The first aspect: Visual Networking management, mainly dealing with the legality of devices in the Visual Networking and communication between devices; for example, the Visual Networking protocol stack of the VRB device and the Visual Networking Agent of the X-VRB device need to enter the Visual Networking to establish services such as video calls.

[0172] Among them, the network access and withdrawal management can implement the legality authentication of Visual Networking devices;

[0173] The keep-alive management is the dynamic detection of Visual Networking devices; the data management is the management of data from the data transmission level, that is, data management according to different data requirements; the service management is the management of each service to ensure the normal operation of various services, such as meetings, video calls, live broadcasts, etc.

[0174] The second aspect: X-VRB management, including Visual Networking network connection management, protocol conversion management, dynamic IP learning management, and log management;

[0175] Visual networking network connection management refers to: The visual networking agent of the X-VRB device has visual networking management functions. For the optimal allocation of network resources, visual networking management layer data (such as network access, heartbeat, control messages, etc.) is managed through the management port (i.e., the corresponding port on the network card of the X-VRB device);

[0176] Protocol conversion management refers to: The protocol conversion unit of the X-VRB device is responsible for the conversion between the RDMA protocol (such as the RoCE v2 protocol) and the visual networking protocol (such as the VRB protocol). The converted data conducts data communication through the service port (i.e., the corresponding port on the network card of the X-VRB device);

[0177] Dynamic IP learning management refers to: The MAC address, IP address, and port number of the access device (i.e., the RDMA device accessing the X-VRB) are quickly matched to achieve fast data forwarding of VRB converted into RoCE v2 data;

[0178] Log management records the operating conditions of the X-VRB device.

[0179] Third aspect: VRB management, including network connection management, dynamic IP learning management, event synchronization management, network quality management, and log management;

[0180] Network connection management refers to: Conducting visual networking management;

[0181] Dynamic IP learning management refers to: After the VRB device receives the data packet sent by the X-VRB device, it extracts the IP address of the RDMA device in the data packet and records it in the queue. Thus, when data needs to be returned to the RDMA device, the IP address of the RDMA device can be extracted from the queue to facilitate the encapsulation and sending of data packets.

[0182] Event synchronization management is responsible for synchronizing data to the upper layer when receiving or sending data;

[0183] Network quality management is responsible for the detection, regulation, and load balancing of network quality during the transmission of the VRB device;

[0184] Log management records the operating conditions of the VRB device.

[0185] In summary, in the multi-protocol interworking and fusion method based on VRB in the embodiments of the present application, the X-VRB device is responsible for establishing a connection and data communication with the VRB device side on the one hand, and communicating with the RDMA device on the other hand. It converts the received RoCE v2 protocol frame of the RDMA into VRB protocol frame data and sends it to the VRB device, and converts the received VRB data into RoCE v2 protocol frame and sends it to the RDMA device, thereby realizing data fusion and synchronization.

[0186] It should also be noted that when performing distributed training in an AI cluster, communication is an essential link and also an additional system overhead compared to single-machine training. The time ratio between communication and computing often determines the upper limit of the acceleration ratio of a distributed machine learning system. Therefore, the key to distributed machine learning is to design a communication mechanism to reduce the time ratio between communication and computing and train a high-precision model more efficiently. The embodiments of the present application can implement the VRB protocol and RoCE v2 protocol interoperability and synchronization technology for large AI models, support multiple protocols for AI models, and solve the problem that VRB cannot be compatible with other RDMA protocols, so that the AI cluster can achieve more efficient communication.

[0187] It should be noted that for the method embodiments, for the sake of simple description, they are all expressed as a series of action combinations. However, those skilled in the art should know that the embodiments of the present application are not limited by the described action sequence, because according to the embodiments of the present application, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily essential for the embodiments of the present application.

[0188] In a third aspect, an embodiment of the present application provides a multi-protocol interoperability and fusion device based on VRB; as Figure 8 shown, applied to a first Visual Networking device, the device may include the following modules:

[0189] A first receiving module 801, configured to receive a first data packet sent by an RDMA device, where the first data packet includes first datagram header information of an RDMA protocol and first data content;

[0190] A first processing module 802, configured to encapsulate a Visual Networking protocol for the first data packet to obtain a second data packet, where the second data packet includes first Visual Networking protocol information, the first datagram header information, and the first data content;

[0191] A first sending module 803, configured to send the second data packet to a second Visual Networking device based on the first Visual Networking protocol information, so that the second Visual Networking device transmits the first data content to a first large model deployed on the second Visual Networking device according to at least part of the information in the first datagram header information.

[0192] In some embodiments of the present application, the first datagram header information includes the IP address of the second Visual Networking device; the device further includes:

[0193] A first obtaining module, configured to obtain the physical address of the second Visual Networking device according to the IP address of the second Visual Networking device;

[0194] The first encapsulation module is used to encapsulate the physical address of the second Visual Networking device into the second data packet;

[0195] The first sending module 803 is specifically configured to: send the second data packet to the second Visual Networking device based on the first Visual Networking protocol information and the physical address of the second Visual Networking device.

[0196] In some embodiments of the present application, the device further includes:

[0197] The third receiving module is used to receive a third data packet sent by the second Visual Networking device based on the second Visual Networking protocol information, where the third data packet includes the second Visual Networking protocol information, the second data header information of the RDMA protocol, and the second data content, and the second data content includes: the content output by the first large model after the first data content is input into the first large model;

[0198] The second processing module is used to remove the second Visual Networking protocol information in the third data packet to obtain a fourth data packet, where the fourth data packet includes the second data header information and the second data content;

[0199] The third sending module is used to send the fourth data packet to the RDMA device based on the second data header information.

[0200] In some embodiments of the present application, the second data header information includes the IP address of the RDMA device; the device further includes:

[0201] The second obtaining module is used to obtain the physical address of the RDMA device according to the IP address of the RDMA device;

[0202] The third processing module is used to encapsulate the physical address of the RDMA device into the fourth data packet;

[0203] The third sending module is specifically configured to: send the fourth data packet to the RDMA device based on the physical address of the RDMA device and the second data header information.

[0204] In some embodiments of the present application, the device further includes:

[0205] The fourth receiving module is used to receive a fifth data packet sent by the second Visual Networking device based on the third Visual Networking protocol information, where the fifth data packet includes the third Visual Networking protocol information, the physical address of the RDMA device, and the third data content;

[0206] The first determination module is configured to determine the IP address of the RDMA device according to the physical address of the RDMA device;

[0207] The fourth processing module is configured to remove the third Visual Networking Protocol information in the fifth data packet and process the fifth data packet according to the IP address of the RDMA device to obtain a sixth data packet, where the sixth data packet includes the physical address of the RDMA device, the third datagram header information of the RDMA protocol, and the third data content, and the third datagram header information includes the IP address of the RDMA device;

[0208] The fourth sending module is configured to send the sixth data packet to the RDMA device based on the physical address of the RDMA device and the third datagram header information, so that the RDMA device transmits the third data content to the second large model deployed on the RDMA device according to at least part of the information in the third datagram header information.

[0209] In a fourth aspect, an embodiment of the present application provides a multi-protocol interworking and fusion device based on VRB, which is applied to a second Visual Networking device, such as Figure 9 As shown, the device may include the following modules:

[0210] The second receiving module 901 is configured to receive a second data packet sent by a first Visual Networking device based on first Visual Networking Protocol information, where the second data packet includes the first Visual Networking Protocol information, the first datagram header information of the RDMA protocol, and the first data content;

[0211] The second sending module 902 is configured to transmit the first data content to the first large model deployed on the second Visual Networking device according to at least part of the information in the first datagram header information.

[0212] In some embodiments of the present application, the first datagram header information includes the IP address of the second Visual Networking device; the second data packet further includes: the physical address of the second Visual Networking device determined according to the IP address of the second Visual Networking device.

[0213] In some embodiments of the present application, the device further includes:

[0214] The fifth processing module is configured to encapsulate the Visual Networking Protocol and the RDMA protocol for the second data content after obtaining the second data content to obtain a third data packet, where the second data content includes: the content output by the first large model after the first data content is input into the first large model, and the third data packet includes second Visual Networking Protocol information, second datagram header information of the RDMA protocol, and the second data content;

[0215] A fifth sending module, configured to send the third data packet to the first Visual Networking Protocol (VNP) device based on the second VNP information, so that the first VNP device removes the second VNP information in the third data packet to obtain a fourth data packet, and sends the fourth data packet to the RDMA device based on the second data packet header, where the fourth data packet includes the second data header information and the second data content.

[0216] In some embodiments of the present application, the first data header information further includes the IP address of the RDMA device; the apparatus further includes:

[0217] An extraction module, configured to extract the IP address of the RDMA device in the first data header information and record it;

[0218] The fifth processing module encapsulates the RDMA protocol for the second data content, specifically:

[0219] According to the IP address of the RDMA device, encapsulate the RDMA protocol for the second data content to obtain the second data header information including the IP address of the RDMA device.

[0220] In some embodiments of the present application, the apparatus further includes: a channel establishment module, configured to establish a connection channel with the first VNP device;

[0221] The extraction module is specifically configured to: when receiving the connection channel of the second data packet, which is the connection channel between the second VNP device and the first VNP device, extract the IP address of the RDMA device in the first data header information.

[0222] In some embodiments of the present application, the apparatus further includes:

[0223] A sixth sending module, configured to send a fifth data packet to the first VNP device based on third VNP information, so that the first VNP device determines the IP address of the RDMA device according to the physical address of the RDMA device, removes the third VNP information in the fifth data packet, processes the fifth data packet according to the IP address of the RDMA device to obtain a sixth data packet, and sends the sixth data packet to the RDMA device based on the physical address of the RDMA device and the third data header information of the RDMA protocol;

[0224] wherein the fifth data packet includes the third VNP information, the physical address of the RDMA device, and third data content;

[0225] The sixth data packet includes the physical address of the RDMA device, the third data header information, and the third data content, and the third data header information includes the IP address of the RDMA device.

[0226] For the apparatus embodiments, since they are basically similar to the method embodiments, the description is relatively simple. For the relevant parts, please refer to the partial description of the method embodiments.

[0227] In a fifth aspect, an embodiment of the present application further provides a VRB-based multi-protocol interworking and fusion system, including a first Visual Networking device and a second Visual Networking device. Among them, the first Visual Networking device is used to execute the VRB-based multi-protocol interworking and fusion method described in the first aspect above, and the second Visual Networking device is used to execute the VRB-based multi-protocol interworking and fusion method described in the second aspect above.

[0228] In a sixth aspect, an embodiment of the present application further provides an electronic device, including:

[0229] One or more processors; and

[0230] One or more machine-readable media storing instructions, which when executed by the one or more processors, cause the electronic device to execute the VRB-based multi-protocol interworking and fusion method described in the first aspect above, or execute the VRB-based multi-protocol interworking and fusion method described in the second aspect above.

[0231] In a seventh aspect, an embodiment of the present application provides a computer-readable storage medium, and the computer program stored therein causes a processor to execute the VRB-based multi-protocol interworking and fusion method described in the first aspect above, or execute the VRB-based multi-protocol interworking and fusion method described in the second aspect above.

[0232] Each embodiment in this specification is described in a progressive manner. Each embodiment focuses on the differences from other embodiments. For the same or similar parts among the embodiments, reference can be made to each other.

[0233] Those skilled in the art should understand that the embodiments of the present application can be provided as methods, apparatuses, or computer program products. Therefore, the embodiments of the present application can take the form of completely hardware embodiments, completely software embodiments, or embodiments combining software and hardware aspects. Moreover, the embodiments of the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memories, CD-ROMs, optical memories, etc.) containing computer-usable program codes.

[0234] Embodiments of the present application are described with reference to the flowcharts and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing terminal devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing terminal devices generate a device for implementing the functions specified in one process Figure 1 one process or multiple processes and / or blocks Figure 1 or multiple blocks.

[0235] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing terminal device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including an instruction device, and the instruction device implements the functions specified in one process Figure 1 one process or multiple processes and / or blocks Figure 1 or multiple blocks.

[0236] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal device, so that a series of operation steps are executed on the computer or other programmable terminal device to generate a computer-implemented process. Thus, the instructions executed on the computer or other programmable terminal device provide steps for implementing the functions specified in one process Figure 1 one process or multiple processes and / or blocks Figure 1 or multiple blocks.

[0237] Although the preferred embodiments of the embodiments of the present application have been described, those skilled in the art can make additional changes and modifications to these embodiments once they learn the basic creative concepts. Therefore, the appended claims are intended to be construed as including the preferred embodiments and all changes and modifications falling within the scope of the embodiments of the present application.

[0238] Finally, it should also be noted that in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or terminal device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or terminal device. Without more limitations, an element defined by the statement "comprising an..." does not exclude the existence of additional identical elements in the process, method, article or terminal device comprising the said element.

[0239] The above has introduced in detail a multi-protocol interoperability and fusion method, apparatus, device and storage medium based on VRB provided by the present application. Specific examples are used in this text to elaborate on the principle and implementation manner of the present application. The description of the above embodiments is only used to help understand the method and its core idea of the present application; at the same time, for those of ordinary skill in the art, according to the idea of the present application, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to the present application.

Claims

1. A multi - protocol interoperability and fusion method based on VRB, characterized in that Applied to the first Visual Networking device, the method includes: Receiving a first data packet sent by a Remote Direct Memory Access (RDMA) device, where the first data packet includes first datagram header information of the RDMA protocol and first data content; Encapsulating the Visual Networking protocol for the first data packet to obtain a second data packet, where the second data packet includes first Visual Networking protocol information, the first datagram header information, and the first data content; Based on the first Visual Networking protocol information, sending the second data packet to a second Visual Networking device, so that the second Visual Networking device transmits the first data content to a first large model deployed on the second Visual Networking device according to at least part of the information in the first datagram header information.

2. The method according to claim 1, wherein The first datagram header information includes the IP address of the second Visual Networking device; The method further includes: Obtaining the physical address of the second Visual Networking device according to the IP address of the second Visual Networking device; Encapsulating the physical address of the second Visual Networking device into the second data packet; The sending the second data packet to the second Visual Networking device based on the first Visual Networking protocol information includes: Based on the first Visual Networking protocol information and the physical address of the second Visual Networking device, sending the second data packet to the second Visual Networking device.

3. The method according to claim 1 or 2, characterized in that, After the first data content is transmitted to the first large model, the method further includes: Receiving a third data packet sent by the second Visual Networking device based on second Visual Networking protocol information, where the third data packet includes the second Visual Networking protocol information, second datagram header information of the RDMA protocol, and second data content, and the second data content includes: the content output by the first large model after the first data content is input into the first large model; Removing the second Visual Networking protocol information from the third data packet to obtain a fourth data packet, where the fourth data packet includes the second datagram header information and the second data content; Based on the second datagram header information, sending the fourth data packet to the RDMA device.

4. The method according to claim 3, characterized in that, The second datagram header information includes the IP address of the RDMA device; The method further includes: Obtaining the physical address of the RDMA device according to the IP address of the RDMA device; Encapsulating the physical address of the RDMA device into the fourth data packet; The sending the fourth data packet to the RDMA device based on the second datagram header information includes: Based on the physical address of the RDMA device and the second datagram header information, sending the fourth data packet to the RDMA device.

5. The method according to claim 1 or 2, characterized in that, The method further includes: Receiving a fifth data packet sent by the second Visual Networking device based on third Visual Networking protocol information, where the fifth data packet includes the third Visual Networking protocol information, the physical address of the RDMA device, and third data content; Determining the IP address of the RDMA device according to the physical address of the RDMA device; Remove the third Visual Networking Protocol information in the fifth data packet, and process the fifth data packet according to the IP address of the RDMA device to obtain a sixth data packet, where the sixth data packet includes the physical address of the RDMA device, the third data header information of the RDMA protocol, and the third data content, and the third data header information includes the IP address of the RDMA device; Based on the physical address of the RDMA device and the third data header information, send the sixth data packet to the RDMA device, so that the RDMA device transmits the third data content to the second large model deployed on the RDMA device according to at least part of the information in the third data header information.

6. A multi-protocol interworking and fusion method based on VRB, characterized in that, Applied to a second Visual Networking device, the method includes: Receive a second data packet sent by a first Visual Networking device based on first Visual Networking Protocol information, where the second data packet includes the first Visual Networking Protocol information, the first data header information of the RDMA protocol, and first data content; Transmit the first data content to the first large model deployed on the second Visual Networking device according to at least part of the information in the first data header information.

7. The method according to claim 6, wherein The first data header information includes the IP address of the second Visual Networking device; The second data packet further includes: the physical address of the second Visual Networking device determined according to the IP address of the second Visual Networking device.

8. The method according to claim 6 or 7, characterized in that The method further includes: After obtaining the second data content, encapsulate the Visual Networking protocol and the RDMA protocol for the second data content to obtain a third data packet, where the second data content includes: the content output by the first large model after the first data content is input into the first large model, and the third data packet includes second Visual Networking Protocol information, the second data header information of the RDMA protocol, and the second data content; Send the third data packet to the first Visual Networking device based on the second Visual Networking Protocol information, so that the first Visual Networking device removes the second Visual Networking Protocol information in the third data packet to obtain a fourth data packet, and sends the fourth data packet to the RDMA device based on the second data packet header, where the fourth data packet includes the second data header information and the second data content.

9. The method according to claim 8, wherein The first data header information further includes the IP address of the RDMA device; The method further includes: Extract and record the IP address of the RDMA device in the first data header information; The encapsulating the RDMA protocol for the second data content includes: According to the IP address of the RDMA device, encapsulate the RDMA protocol for the second data content to obtain the second data header information including the IP address of the RDMA device.

10. The method according to claim 9, wherein The method further includes: Establish a connection channel with the first Visual Networking device; The extracting the IP address of the RDMA device in the first data header information includes: When the connection channel for receiving the second data packet is the connection channel between the second Visual Networking device and the first Visual Networking device, extract the IP address of the RDMA device in the first data header information.

11. The method according to claim 6 or 7, characterized in that The method further includes: Sending a fifth data packet to the first Visual Networking device based on the third Visual Networking protocol information, so that the first Visual Networking device determines the IP address of the RDMA device according to the physical address of the RDMA device, removes the third Visual Networking protocol information in the fifth data packet, processes the fifth data packet according to the IP address of the RDMA device to obtain a sixth data packet, and sends the sixth data packet to the RDMA device based on the physical address of the RDMA device and the third data header information of the RDMA protocol; Wherein, the fifth data packet includes the third Visual Networking protocol information, the physical address of the RDMA device, and third data content; The sixth data packet includes the physical address of the RDMA device, the third data header information, and the third data content, and the third data header information includes the IP address of the RDMA device.

12. A multi - protocol interoperability and fusion device based on VRB, characterized in that, Applied to a first Visual Networking device, the apparatus includes: A first receiving module, configured to receive a first data packet sent by an RDMA device, where the first data packet includes first data header information of the RDMA protocol and first data content; A first processing module, configured to encapsulate a Visual Networking protocol for the first data packet to obtain a second data packet, where the second data packet includes first Visual Networking protocol information, the first data header information, and the first data content; A first sending module, configured to send the second data packet to a second Visual Networking device based on the first Visual Networking protocol information, so that the second Visual Networking device transmits the first data content to a first large model deployed on the second Visual Networking device according to at least part of the information in the first data header information.

13. A multi - protocol interoperability and fusion device based on VRB, characterized in that, Applied to a second Visual Networking device, the apparatus includes: A second receiving module, configured to receive a second data packet sent by a first Visual Networking device based on first Visual Networking protocol information, where the second data packet includes the first Visual Networking protocol information, first data header information of the RDMA protocol, and first data content; A second sending module, configured to transmit the first data content to a first large model deployed on the second Visual Networking device according to at least part of the information in the first data header information.

14. An electronic device, characterized in that, Includes: One or more processors; And One or more machine-readable media storing instructions that, when executed by the one or more processors, cause the electronic device to execute the VRB-based multi-protocol interworking and fusion method according to any one of claims 1 to 5, or execute the VRB-based multi-protocol interworking and fusion method according to any one of claims 6 to 11.

15. A computer-readable storage medium, characterized in that, The stored computer program causes the processor to execute the VRB-based multi-protocol interworking and fusion method according to any one of claims 1 to 5, or execute the VRB-based multi-protocol interworking and fusion method according to any one of claims 6 to 11.

Images