DPU-oriented lossless data center exchange networking method and device

By designing the RDMA switching matrix of controller and fully connected matrix in the DPU data center, using optimal path calculation and address encoding, the problem of insufficient performance of traditional RoCE protocols is solved, and high-performance lossless data exchange and flexible expansion between DPUs is achieved. , improves the communication performance and expansion capabilities of the data center.

CN119011464APending Publication Date: 2024-11-22SUZHOU HENGTIAN INFORMATION TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411241019.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2024-08-29
Filing Date
2024-09-05
Publication Date
2024-11-22

AI Technical Summary

Technical Problem

The existing DPU data center networking method, especially the RoCE protocol based on traditional Ethernet, cannot achieve the performance of the InfiniBand network, and there are problems of network delay and link damage, and insufficient scalability and openness.

Method used

Design a lossless data center switching networking method for DPU. Through the RDMA switching matrix of the controller and the fully connected matrix, the lossless RDMA data exchange based on the RoCE protocol is realized, and the optimal path calculation and address encoding are used to build a high-performance DPU. intercommunication network and supports horizontal and vertical expansion.

Benefits of technology

It realizes high-performance RDMA switch networking without DPU perception, avoids RDMA performance losses caused by traditional L2/L3 switch protocols, supports flexible expansion and rapid capacity expansion, and realizes transmission performance of approximate point-to-point networking.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119011464A_ABST
    Figure CN119011464A_ABST
Patent Text Reader

Abstract

The invention discloses a DPU (Data Processing Unit)-oriented lossless data center exchange networking method and device, belongs to the technical field of data processing units, and aims at realizing efficient lossless communication by calculating an optimal path of a controller, controlling the on-off of a fully-connected matrix and constructing a lossless high-speed network between DPUs. Transmission performance of high-performance RDMA switching networking and approximate point-to-point networking without perception of the DPU is realized, and RDMA transmission performance loss caused by a traditional L2 / L3 switching protocol is avoided; meanwhile, through an RDMA address coding and addressing method of the RDMA switching matrix, an existing RoCE protocol is supported, and DPU non-perception data switching is achieved; in addition, flexible transverse and longitudinal expansion is supported through an RDMA switching matrix expansion method, and rapid expansion of the system is realized through superposition of a plurality of devices.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of data processing units, and in particular to a lossless data center switching networking method and apparatus for DPUs. Background Technology

[0002] A Data Processing Unit (DPU) is a programmable processor built around a data-centric architecture, designed to efficiently handle large-scale data transmission tasks within data centers, including data transmission, reduction, protection, compression, analysis, and encryption. With the rise of technologies such as large AI models, increasingly higher demands are being placed on the transmission of training and inference data within data centers, leading to increasingly complex DPU networking within these centers.

[0003] To meet the communication networking requirements of DPUs, data center DPU networking primarily employs two methods: InfiniBand networks and RoCE (RDMA over Converged Ethernet) networks. InfiniBand networks create a dedicated, protected channel directly between nodes via InfiniBand switches and manage and execute remote data access through InfiniBand network interface cards (NICs). However, they rely on proprietary protocols and devices, resulting in poor openness and scalability. In contrast, RoCE networks use the open RoCE protocol, overcoming the limitations of InfiniBand and allowing for use in traditional Ethernet. However, traditional Ethernet is not optimized for the RoCE protocol and cannot match the performance of InfiniBand networks in terms of network latency, link impairments, and topology scalability. Summary of the Invention

[0004] This invention proposes a lossless data center switching networking method and apparatus for DPUs. By designing an RDMA switching matrix including a controller and a fully connected matrix, lossless RDMA data switching based on the RoCE protocol is realized. This can effectively avoid the loss of the RoCE protocol on traditional Ethernet and realize high-performance RDMA switching networking that is imperceptible to DPUs.

[0005] A lossless data center switching networking method and apparatus for DPUs.

[0006] Includes the following steps:

[0007] S1. When a DPU device is connected, the controller saves the mapping between the DPU device's MAC address and the access port code, and updates the exchange table.

[0008] S2. When the accessed DPU device initiates a message exchange request, the controller parses the ether_header in the message header to obtain the destination MAC address and source MAC address, and calculates the optimal path.

[0009] S3. The controller controls the connectivity of points along the optimal path, establishes communication connections, and enables message exchange.

[0010] S4. After message exchange is completed, the controller closes the connectivity point and stops message exchange.

[0011] When multiple controllers exist, the device with the smallest management port MAC address is the master controller, and the other devices are slave controllers. The master controller receives the exchange table and topology status sent periodically by all slave controllers.

[0012] If the DPU device is connected to the slave control device, the master control device receives the updated exchange table sent by the slave control device.

[0013] If a DPU device sends a message exchange request and the receiving device is a slave control device, the master control device receives the parsed MAC address sent by the slave control device, performs optimal path calculation, and sends control commands to the slave control devices. All control devices control the connection points in their respective optimal paths to establish communication connections and perform message exchange between DPU devices.

[0014] The optimal path in S1 is calculated as follows:

[0015] Let D i,j,k Let i be the shortest path length from i to j, using only connected nodes in the set (1..k) as intermediate nodes;

[0016] If the shortest path passes through connected vertex k, then D i,j,k =D i,k,k-1 +D k,j,k-1 ;

[0017] If the shortest path does not pass through connected vertex k, then D i,j,k =D i,k,k-1 ;

[0018] Therefore, D i,j,k =min(D i,j,k-1 D i,k,k-1 + k,j,k-1 );

[0019] If the connected point k is in a conducting state, and backoff is executed, then D i,j,k =D i,k,k-1 .

[0020] In S1, the controller uniformly encodes all ports of the device using 16 bits, with the high 4 bits representing the row number of the port in the fully connected matrix and the low 12 bits representing the column number of the port in the fully connected matrix.

[0021] When the DPU device is connected, the controller adds the mapping between the MAC address and the access port code to the switching table.

[0022] In S2, the controller parses the ether_header in the packet header, looks up the destination port number based on the updated exchange table, and plans the optimal path based on the source port number and the destination port number.

[0023] A lossless data center switching networking device for DPUs includes a controller, a fully connected matrix, and an external interface;

[0024] The external interface includes a left expansion terminal, a right expansion terminal, an SPF port, and a management network port;

[0025] The fully connected matrix includes M columns and N rows of connected nodes, forming an M*N connected matrix. Each connected node has four endpoints, and the connected nodes can control any two endpoints to be connected, thereby realizing data exchange.

[0026] Both the left extension terminal and the right extension terminal contain multiple connection points, which are connected by cables.

[0027] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention uses the controller to calculate the optimal path and control the connection and disconnection of the fully connected matrix to construct a lossless high-speed network between DPUs, achieving efficient and lossless communication. It enables DPUs to achieve high-performance RDMA switching networking and near-point-to-point networking transmission performance without being aware of each other, avoiding the RDMA transmission performance loss caused by traditional L2 / L3 switching protocols. At the same time, through the RDMA address encoding and addressing method of the RDMA switching matrix, it supports the existing RoCE protocol and realizes DPU-invisible data exchange. In addition, it supports flexible horizontal and vertical expansion through the RDMA switching matrix expansion method, and achieves rapid system expansion through the superposition of multiple devices. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the lossless switching device of the present invention;

[0029] Figure 2 This is a schematic diagram of the fully interconnected matrix of the lossless switching equipment of the present invention;

[0030] Figure 3 This is a schematic representation of the lossless switching device of the present invention.

[0031] Figure 4 This is a schematic diagram of the horizontal and vertical expansion connection of the lossless switching device of the present invention;

[0032] Figure 5 This is a schematic diagram of the deployment of the lossless switching device of the present invention in a data center. Detailed Implementation

[0033] A specific embodiment of the present invention is described in detail below with reference to the accompanying drawings, but it should be understood that the protection scope of the present invention is not limited by the specific embodiment.

[0034] Example 1

[0035] like Figures 1-4 As shown, this embodiment of the invention proposes a lossless data center switching networking device for DPUs. The lossless L2 switching equipment is designed for multi-DPU networking in data centers, and includes a controller, a fully connected matrix, and external interfaces.

[0036] The controller is used for device control, including controlling the state of connected points in the full connectivity matrix, storing the exchange table in the device, planning message exchange paths, controlling the horizontal and vertical expansion of the device, and other device operation control;

[0037] A fully connected matrix is ​​used to achieve lossless message exchange. It consists of connected nodes, the lines between connected nodes, and the lines between connected nodes and the controller.

[0038] like Figure 2 As shown, the fully connected matrix contains M columns and N rows of connected nodes, forming an M*N connected matrix that can support a maximum of 2*M-2*N-4 DPU devices. At the same time, each connected node has four endpoints, and the connected node can control any two endpoints to conduct, realizing data exchange. The controller is used to control the connectivity of the connected nodes in the device according to the data exchange requirements of the DPU connected to the device, so that the exchange matrix can meet different data exchange requirements.

[0039] The external interfaces are used for external connections and expansion, including the left expansion terminal, right expansion terminal, SPF port, and management network port;

[0040] Both the left and right expansion terminals have N connection points for cable connection. The top and bottom sides of the device have M SFP connection terminals for fiber optic connection. The right side of the device has an RJ45 management and control port, which supports connection to a computer host for management and operation via Ethernet cable.

[0041] like Figure 4 As shown, when expanding the equipment horizontally, the left and right expansion terminals of multiple devices are interconnected by cables to form an M*N*i fully connected matrix, thereby achieving horizontal expansion, where i is the number of devices;

[0042] When expanding the equipment vertically, the upper and lower terminals of multiple devices are interconnected using optical fibers to form a j*M*N connected matrix, thereby achieving vertical expansion, where j is the number of devices;

[0043] Correspondingly, by connecting the left and right and top and bottom terminals of multiple devices to each other, a j*M*N*i connected matrix can be formed, enabling simultaneous horizontal and vertical expansion, where i and j are the number of devices, and a total of i*j devices are required.

[0044] During expansion, the management ports of all devices are connected through switches or routers to enable communication between the device control planes.

[0045] Example 2

[0046] This invention provides a lossless data center switching network method for DPUs, comprising the following steps:

[0047] S1. When a DPU device is connected, the controller saves the mapping between the DPU device's MAC address and the access port code, and updates the exchange table.

[0048] S2. When the accessed DPU device initiates a message exchange request, the controller parses the ether_header in the message header to obtain the destination MAC address and source MAC address, and calculates the optimal path.

[0049] S3. The controller controls the connectivity of points along the optimal path, establishes communication connections, and enables message exchange.

[0050] S4. After message exchange is completed, the controller closes the connectivity point and stops message exchange;

[0051] The RDMA address encoding and forwarding table design method for the RDMA switching matrix in this embodiment is as follows:

[0052] The controller encodes all ports of the device using a 16-bit encoding. The high 4 bits represent the row number of the port in the fully connected matrix, and the low 12 bits represent the column number of the port in the fully connected matrix. Correspondingly, each row of the forwarding table contains 72 bits, of which the high 48 bits are the MAC address, the middle 16 bits are the port code, and the low 8 bits are reserved for extensions. When the DPU device's network port is connected to the device, the controller adds the mapping between the MAC address and the port code to the switching table.

[0053] The RDMA addressing and forwarding method for the RDMA switching matrix in this embodiment is as follows:

[0054] During addressing, the controller first parses the ether_header field in the RoCE message header, and then finds the destination port number by looking up the exchange table. During forwarding, the controller uses the optimal path calculation algorithm to plan the path based on the source port number and the destination port number, controls the connectivity of the connected points in the path, and connects the corresponding paths in the connectivity matrix to realize message forwarding between DPUs. After forwarding is completed, the controller closes the connectivity matrix on the path.

[0055] Lossless switching equipment achieves vertical expansion through SPF port connections and horizontal expansion through left and right expansion terminals. During expansion, the management port of the equipment is connected to the management switch to enable communication between controllers. The expansion steps are as follows:

[0056] S11. After the equipment is expanded, the equipment controller communicates through the management network port. The device with the smallest MAC address of the management network port is the master control device, and the other devices are slave control devices.

[0057] S12. All slave control devices periodically send their exchange tables and topology status to the master control device. The master control device saves the global exchange table and global topology status.

[0058] When S13 and DPU devices are connected, the controller updates the exchange table and saves the mapping between the device's MAC address and the access port code. If the connected device is a slave control device, the updated exchange table is sent to the master control device.

[0059] S14. When the DPU device initiates a message exchange request, the controller parses the ether_header in the message header to obtain the destination MAC address and the source MAC address.

[0060] When the S15 and DPU devices initiate a message exchange request, the two communicating parties will establish a communication connection. If the receiving device is the slave control device, it will send the MAC address parsed in the previous step to the master control device for path calculation.

[0061] S16. When the two communicating parties establish a communication connection, if the main control device receives the message exchange request, the calculation will start directly.

[0062] S17. The main control device calculates the optimal path according to the algorithm and sends the control command to the controller of the slave control device.

[0063] S18. All controllers control the connection of the connection points in their respective optimal paths, complete the establishment of communication connections, and realize message exchange between DPU devices.

[0064] S19. After message exchange is completed, all controllers close the connectivity points in the path and stop message exchange;

[0065] In this embodiment, the device connects to the DPU device's network port via an Ethernet port to achieve lossless message exchange between the DPU devices, such as... Figure 5 As shown;

[0066] The calculation method for the above optimal path is as follows:

[0067] Let D i,j,kLet i be the shortest path length from i to j, using only connected nodes in the set (1..k) as intermediate nodes;

[0068] If the shortest path passes through connected vertex k, then D i,j,k =D i,k,k-1 +D k,j,k-1 ;

[0069] If the shortest path does not pass through connected vertex k, then D i,j,k =D i,k,k-1 ;

[0070] Therefore, D i,j,k =min(D i,j,k-1 D i,k,k-1 +D k,j,k-1 );

[0071] If the connected point k is in a conducting state, and backoff is executed, then D i,j,k =D i,k,k-1 .

[0072] In S12 above, the slave control device periodically sends topology information to the master control device. The topology information includes two parts: the MAC addresses of the slave device's upper, lower, left, and right devices, and the corresponding unified 16-bit encoding of the slave device's internal ports. The master control device calculates the global topology using a depth-first search based on the status of the upper, lower, left, and right devices. The global topology includes two parts: the connection relationship of all devices and the port encoding of each device's internal connection point. In the global topology, the master control device encodes all devices using 8 bits.

[0073] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit and essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0074] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A DPU-oriented lossless data center switching networking method, characterized in that: The steps include: S1. When a DPU device is connected, the controller saves the mapping between the DPU device MAC address and the access port code, and updates the switching table; S2: When the connected DPU device initiates a message exchange request, the controller parses the ether_header in the message header, obtains the destination MAC address and source MAC address, and calculates the optimal path; S3: The controller controls the connectivity of the connectivity points in the optimal path, establishes a communication connection, and implements message exchange S4. After the message exchange is completed, the controller closes the connection point and stops the message exchange.

2. The DPU-oriented lossless data center switching networking method according to claim 1, characterized in that: When there are multiple controllers, the device with the smallest management port MAC address is the master control device, and the other devices are slave control devices. The master control device receives the switching table and topology status sent regularly by all slave control devices.

3. The DPU-oriented lossless data center switching networking method according to claim 2, characterized in that: If the DPU device is connected to the slave control device, the master control device receives the updated switching table sent by the slave control device.

4. The DPU-oriented lossless data center switching networking method according to claim 2, characterized in that: If the DPU device sends a message exchange request and the device that receives the request is a slave control device, the master control device receives the resolved MAC address sent by the slave control device and performs optimal path calculation. The master control device sends the control instruction to the slave control device, and all control devices control the connection points involved in the optimal path to be connected, establish a communication connection, and exchange messages between DPU devices.

5. The DPU-oriented lossless data center switching networking method according to claim 1 or 4, characterized in that: The optimal path calculation method in S1 is as follows: Assume D i,j,k is the shortest path length from D to j, with only connected points in the (1..k) set as intermediate nodes; If the shortest path passes through the connected point k, then D i,j,k =D i,k,k-1 +D k,j,k-1 ; If the shortest path does not pass through the connected point k, then D i,j,k =D i,k,k-1 ; Therefore, D i,j,k =min(D i,j,k-1 ,D i,k,k-1 +D k,j,k-1 ); If the connection point k is in the conducting state and backoff is executed, then D i,j,k =D i,k,k-1 .

6. The DPU-oriented lossless data center switching networking method according to claim 1, characterized in that: In S1, the controller uniformly encodes all ports of the device with 16 bits, where the upper 4 bits are the row number of the port in the fully connected matrix, and the lower 12 bits are the column number of the port in the fully connected matrix; When a DPU device is connected, the controller adds the mapping between the MAC address and the access port code to the switching table.

7. The DPU-oriented lossless data center switching networking method according to claim 1, characterized in that: In S2, the controller parses the ether_header in the message header, finds the destination port number based on the updated switching table, and plans the optimal path according to the source port number and the destination port number.

8. A lossless data center switching networking device for DPU, characterized in that: Including controller, fully connected matrix and external interface; Wherein, the external interface includes a left expansion terminal, a right expansion terminal, an SPF port and a management network port; The fully connected matrix includes M columns and N rows of connected points, forming an M*N connected matrix, each connected point has four endpoints, and the connected point can control any two endpoints to conduct to achieve data exchange.

9. The DPU-oriented lossless data center switching networking device according to claim 8, characterized in that: The left extension terminal and the right extension terminal each include a plurality of connection points, and the connection points are connected by cables.