Data processing system and method and connection equipment

By introducing connected devices into the data processing system, managing the memory address information of multiple computing devices, using the ultra-low-latency communication protocol to interact with the RDMA protocol, the adaptation problem of ultra-low-latency communication protocol and RDMA protocol is solved, and cross-computing clusters are realized Efficiency and flexibility of memory access.

CN118939597AActive Publication Date: 2024-11-12HUAWEI TECH CO LTD
View PDF 8 Cites 0 Cited by

Patent Information

Application Number
CN202411002660.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2022-12-01
Filing Date
2023-03-07
Publication Date
2024-11-12
Estimated Expiration
2043-03-07

AI Technical Summary

Technical Problem

How to adapt the ultra-low-latency communication protocol to the RDMA protocol to solve the memory access latency problem across compute clusters.

Method used

By introducing connected devices into the data processing system, managing the memory address information provided by multiple computing devices, using ultra-low-latency communication protocols (such as CXL protocol or UB protocol) to interact with the RDMA Network Interface Controller (RNIC) to achieve cross-computing Memory access of the cluster, adapted to ultra-low latency communication protocol and RDMA protocol.

Benefits of technology

It improves memory access speed across compute clusters, reduces access latency, enhances memory access flexibility and memory management capabilities between multi-computing clusters.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118939597A_ABST
    Figure CN118939597A_ABST
Patent Text Reader

Abstract

A data processing system, method, and connection device for adapting an ultra-low latency communication protocol, such as a CXL protocol or a UB protocol, to an RDMA protocol implementation. In the application, the data processing system comprises a first computing cluster, a second computing cluster and a connecting device, the first computing cluster comprises a first computing device, and the second computing cluster comprises a second computing device. Furthermore, the connection equipment is connected with the first computing equipment, and the connection equipment is used for managing memory address information of the first computing cluster provided by the first computing equipment and memory address information of the second computing cluster provided by the second computing equipment; the first computing device is used for receiving an access request, wherein the access request is used for accessing the memory space of the second computing device; searching an address of a memory space from memory address information of a second computing cluster managed by the connection device; and accessing the memory space in the second computing device according to the address.
Need to check novelty before this filing date? Find Prior Art

Description

[0001] This application is a divisional application. The application number of the original application is 202310248707.3, and the original application date is March 7, 2023. The entire contents of the original application are incorporated into this application by reference.

[0002] This application claims priority to the Chinese patent application filed with the Intellectual Property Office of the People's Republic of China on December 1, 2022, with application number 202211531954.6 and invention name "A Memory Management Method", the entire contents of which are incorporated by reference in this application. Technical Field

[0003] The present application relates to the field of computing technology, and in particular to a data processing system, method and connection device. Background Art

[0004] Remote direct memory access (RDMA) is conceptually relative to direct memory access (DMA). In DMA technology, external devices (i.e., peripheral component interconnect express (PCIe) devices) can bypass the central processing unit (CPU) to directly access the host's memory; while in RDMA technology, external devices can also bypass the CPU to access the memory on another remote host.

[0005] With the advent of the big data era, data centers not only have to deal with massive amounts of data and users, but also need to provide users with ultra-fast data communication services. The emergence of emerging ultra-low latency communication protocols (such as open interconnect (Compute ExpressLink, CXL)) has brought new possibilities for data communication. Ultra-low latency communication protocols can achieve high-speed and efficient interconnection between CPU and graphics processing unit (GPU), CPU and field programmable gate array (FPGA), or CPU and other accelerators.

[0006] However, how to adapt the ultra-low latency communication protocol to the RDMA protocol is a technical problem that needs to be solved urgently. Summary of the invention

[0007] The present application provides a data processing system, method and connection device for adapting an ultra-low latency communication protocol (such as a CXL protocol or a unified bus (UB) protocol) to an RDMA protocol.

[0008] In a first aspect, the present application provides a data processing system, which includes: a first computing cluster, a second computing cluster and a connection device, wherein the first computing cluster includes a first computing device, and the second computing cluster includes a second computing device. Further, the connection device is connected to the first computing device, and the connection device is used to manage the memory address information of the first computing cluster provided by the first computing device, and the memory address information of the second computing cluster provided by the second computing device; the first computing device is used to: receive an access request, and the access request is used to access the memory space of the second computing device; find the address of the memory space from the memory address information of the second computing cluster managed by the connection device; access the memory space in the second computing device according to the address. In a possible implementation, the data processing system also includes a network device, and the network device is used to connect the first computing cluster and the second computing cluster, and the network device is, for example, an RDMA network interface controller (RNIC). In a possible implementation, the connection device is connected to the first computing device via an ultra-low latency communication protocol (such as a CXL protocol or a UB protocol).

[0009] In the above technical solution, the connecting device manages the memory address information of the computing cluster to which the computing device belongs, which is provided by multiple computing devices respectively. When a computing device connected to the connecting device needs to access the memory space in the computing devices in other computing clusters across computing clusters, the address used to access the memory space in the computing devices in the other computing clusters can be first searched from the connecting device, and then accessed according to the address. In this way, the access based on the RDMA protocol across computing clusters is adapted to the access based on the ultra-low latency communication protocol within the computing cluster.

[0010] In one possible implementation, the first computing cluster also includes a third computing device, and the connecting device is used to connect the first computing device and the third computing device; the first computing device is also used to: obtain memory address information of the third computing device; address the memory address information of the first computing cluster according to the memory address information of the third computing device and the memory address information of the first computing device; and send the memory address information of the first computing cluster to the connecting device.

[0011] In the above technical solution, the first computing device can obtain the memory address information of other computing devices in the first computing cluster to which it belongs, and perform addressing according to the memory address information of other computing devices in the first computing cluster to which it belongs and the memory address information of the first computing device, so as to form the memory address information of the first computing cluster from the perspective of the first computing device. In this way, the first computing device can access the memory space in the first computing cluster based on the ultra-low latency communication protocol, which helps to improve the access speed, that is, reduce the access delay.

[0012] In a possible implementation, the memory address information of the first computing device includes information of multiple memories in the first computing device; the memory address information of the third computing device includes information of multiple memories in the third computing device. In the above technical solution, each computing device can include multiple types of memory, which helps to improve the flexibility of memory access.

[0013] In a possible implementation, the connection device is further used to: receive memory address information of a first computing cluster from a first computing device, and / or receive memory address information of a second computing cluster from a second computing device. In the above technical solution, the connection device can receive memory address information in a computing cluster to which the computing device belongs from a computing device, and then uniformly manage memory address information from multiple computing devices, so as to enable a computing device connected to the connection device to access memory space in computing devices in other computing clusters across computing clusters.

[0014] In one possible implementation, the memory address information of the first computing cluster provided by the first computing device is specifically obtained by the first computing device addressing the memory of the first computing cluster from the operating system level, and the memory address information of the second computing cluster provided by the second computing device is specifically obtained by the second computing device addressing the memory of the second computing cluster from the operating system level. Furthermore, the connection device is also used to further address the sum of the memories of the first computing cluster and the second computing cluster from the application level, obtain the memory address information at the application level, and establish a mapping relationship between the memory address information at the application level and the memory address information of each computing device at the operating system level. In this way, the data processing system can provide the lengths of the memories of multiple computing clusters together to the application, so that the application can access any memory space in the multiple computing clusters, thereby improving the flexibility of application access.

[0015] In one possible implementation, the access request received by the first computing device specifically includes the first address issued by the application. When the first computing device searches for the address of the memory space from the memory address information of the second computing cluster managed by the connection device, it is specifically used as follows: the first computing device sends a query request to the connection device, and the query request includes the first address. Correspondingly, the first computing device receives a query response from the connection device, and the query response includes the second address, wherein the second address is determined by a mapping relationship between the connection device according to the memory address information at the application level and the memory address information of the second computing cluster obtained by addressing the second computing device at the operating system level, and the first address.

[0016] In a possible implementation, the first computing device is further used to determine that the first address accessed by the access request is not in the first computing cluster, that is, to determine that the first address does not belong to a memory address in the first computing cluster. Exemplarily, the first computing device obtains memory address information corresponding to the first computing cluster at the application level from the connection device, and determines that the first address does not belong to a memory address in the first computing cluster based on the first address and the memory address information corresponding to the first computing cluster at the application level. Further, the first computing device also determines that the first address is a memory address in a second computing device belonging to the second computing cluster, that is, to determine that the access request is an access request for accessing the memory space of the second computing device.

[0017] In addition, in another possible implementation, the first computing device is further used to determine that the first address is in the first computing cluster, that is, to determine that the first address belongs to a memory address in the first computing cluster, and then access the memory space indicated by the first address in the first computing cluster. Exemplarily, the first computing device obtains the memory address information corresponding to the first computing cluster at the application level from the connection device, and determines that the first address belongs to a memory address in the first computing cluster according to the first address and the memory address information corresponding to the first computing cluster at the application level.

[0018] In a second aspect, the present application provides a connection device, the device comprising: a transceiver module, used to receive a query request sent by a first computing device, the query request being used to query the address accessed by the access request of the address space of a second computing device of a second computing cluster to be accessed; a management module, used to send the address obtained by the query to the first computing device through the transceiver module based on the query request and the memory address information of the second computing cluster; wherein the connection device is connected to the first computing device in the first computing cluster, and the management module is used to manage the memory address information of the first computing cluster provided by the first computing device, and the memory address information of the second computing cluster provided by the second computing device.

[0019] In a possible implementation, the transceiver module is further configured to receive memory address information of a first computing cluster from a first computing device, and / or receive memory address information of a second computing cluster from a second computing device.

[0020] In a third aspect, the present application provides a computing device, which is specifically a first computing device, and the computing device includes: a transceiver module, used to receive an access request, the access request is used to access the memory space of a second computing device, the first computing device is located in a first computing cluster, and the second computing device is located in a second computing cluster; a search module, used to search for the address of the memory space from a connecting device; an access module, used to access the memory space in the second computing device according to the address; wherein the connecting device is connected to the first computing device, and the connecting device is used to manage the memory address information of the first computing cluster provided by the first computing device, and the memory address information of the second computing cluster provided by the second computing device.

[0021] In a possible implementation, the first computing cluster is connected to the second computing cluster via a network device.

[0022] In a possible implementation manner, the connection device is connected to the first computing device via the CXL protocol or the UB protocol.

[0023] In one possible implementation, the first computing cluster also includes a third computing device, and the connecting device is used to connect the first computing device and the third computing device; the processing module is also used to obtain memory address information of the third computing device; the memory address information of the first computing cluster is formed according to the memory address information of the third computing device and the memory address information of the first computing device; the transceiver module is also used to send the memory address information of the first computing cluster to the connecting device.

[0024] In a possible implementation, the memory address information of the first computing device includes information of multiple memories in the first computing device; and the memory address information of the third computing device includes information of multiple memories in the third computing device.

[0025] In one possible implementation, the processing module is further used to determine, after the transceiver module receives the access request, that the access request is an access request for accessing the memory space of the second computing device based on the fact that the address accessed by the access request (i.e., the first address) does not belong to the memory address of the first computing cluster.

[0026] In a fourth aspect, the present application provides a data processing method, including: a first computing device receives an access request, the access request is used to access the memory space of a second computing device in a second computing cluster; a query request is sent to a connection device, the query request is used to query the address accessed by the access request for accessing the address space of the second computing device, and the first computing device is located in the first computing cluster; the connection device receives the query request, and sends the address obtained by the query to the first computing device based on the query request and the memory address information of the second computing cluster; the first computing device accesses the memory space in the second computing device based on the received address.

[0027] In a possible implementation, the connection device further manages the memory address information of the first computing cluster provided by the first computing device and the memory address information of the second computing cluster provided by the second computing device.

[0028] In a possible implementation manner, the connection device is connected to the first computing device via the CXL protocol or the UB protocol.

[0029] In one possible implementation, the first computing cluster also includes a third computing device, and the connecting device is used to connect the first computing device and the third computing device; the first computing device also obtains memory address information of the third computing device; the first computing device addresses the memory address information of the first computing cluster according to the memory address information of the third computing device and the memory address information of the first computing device; and the first computing device sends the memory address information of the first computing cluster to the connecting device.

[0030] In a possible implementation, the memory address information of the first computing device includes information of multiple memories in the first computing device; and the memory address information of the third computing device includes information of multiple memories in the third computing device.

[0031] In a possible implementation, after receiving the access request, the first computing device determines that the access request is an access request for accessing the memory space of the second computing device based on the fact that the address accessed by the access request does not belong to the memory address of the first computing cluster.

[0032] In a possible implementation, the connection device further receives memory address information of a first computing cluster from the first computing device, and / or receives memory address information of a second computing cluster from the second computing device.

[0033] In a fifth aspect, the present application provides a computer-readable storage medium, in which a computer program or instruction is stored. When the computer program or instruction is executed by a device, the method in any possible implementation manner of the fourth aspect above is implemented.

[0034] In a sixth aspect, the present application provides a computer program product, which includes a computer program or instructions. When the computer program or instructions are executed by a device, the method in any possible implementation manner of the fourth aspect above is implemented.

[0035] In a seventh aspect, the present application provides a device, including a processor, the processor being connected to a memory, the memory being used to store a computer program, and the processor being used to execute the computer program stored in the memory, so that the device implements a method in any possible implementation manner of the above-mentioned fourth aspect.

[0036] The technical effects that can be achieved in any of the second to seventh aspects mentioned above can refer to the description of the beneficial effects in the first aspect mentioned above, and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 A schematic diagram of a host transmitting data to a device via a PCIe protocol provided in this application;

[0038] Figure 2 A schematic diagram of the structure of a data processing system provided for this application;

[0039] Figure 3 A schematic diagram of a memory pool of a data processing system provided by the present application;

[0040] Figure 4 A flowchart of a data processing method provided in this application;

[0041] Figure 5 A schematic diagram of a ring synchronization method provided by the present application;

[0042] Figure 6 A flowchart of another data processing method provided in this application;

[0043] Figure 7 A schematic diagram of the structure of a data processing device provided in this application. DETAILED DESCRIPTION

[0044] The embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0045] In order to facilitate the explanation of the embodiments of the present application, the professional terms and technologies involved in the embodiments of the application are first explained.

[0046] 1. High-speed serial computer expansion bus standard (Peripheral Component Interconnect Express, PCIe) protocol, which belongs to high-speed serial point-to-point dual-channel high-bandwidth transmission. Figure 1The schematic diagram is a host transmitting data to a device via the PCIe protocol. Specifically, the central processing unit (CPU) in the host needs to first write the data to the host's local memory, and then set a flag, where the flag is used to notify the device that the host has written the data to the host's local memory. Accordingly, the device queries the flag, and determines based on the flag that the host has written the data to the host's local memory, and then the device reads the data from the host's local memory and stores the read data in the device's local memory. The way the device transmits data to the host via the PCIe protocol is similar to the way the host transmits data to the device via the PCIe protocol, and will not be repeated here.

[0047] 2. Compute Express Link (CXL) protocol. The CXL protocol is a new protocol based on the PCIe protocol and optimized for cache and memory. The CXL protocol runs on the PCIe physical layer.

[0048] The CXL protocol can be divided into the following three protocols:

[0049] CXL.io protocol is an enumeration configuration protocol, which is mainly used for device discovery and enumeration, error reporting, etc.

[0050] The CXL.cache protocol enables devices to solve memory consistency and access host memory with low latency. Specifically, the CXL.cache protocol mainly provides the ability to quickly access host memory by maintaining cache consistency on the device side. This protocol allows devices to participate in the consistent cache protocol of the CPU in the host. Devices can directly use host memory and the host's local cache, and the host can also obtain data from the device's cache without using memory as an intermediary.

[0051] The CXL.mem protocol allows the host to access device memory as if it were its own local memory. In the CXL.mem protocol, the CPU sends requests to the device, and the device responds to the CPU. The requests sent by the CPU are divided into requests with data and requests without data. Correspondingly, the responses sent by the device are also divided into responses with data and responses without data.

[0052] 3. The unified bus (UB) protocol is an ultra-low latency communication protocol developed by Huawei. It is used to provide high-performance interconnection in data centers, making the interconnection of data centers like a computer.

[0053] 4. Memory pool: The program can apply for a large block of memory from the system in advance to make a memory pool. After that, the program's memory allocation and release are all done in the memory pool. When the memory pool is insufficient, the program will apply for new memory from the system.

[0054] 5. Load / store instructions belong to ARM (advanced risc machine) instructions, which are used to transfer data between registers and memory. Specifically, the load instruction is used to load data from memory into registers; the store instruction is used to store data from registers into memory. Since other ARM instructions can only operate registers, when other ARM instructions need to operate data, they need to first load data from memory to registers through load instructions. Correspondingly, the data executed by other ARM instructions needs to be stored from registers to memory through store instructions.

[0055] Combined with the above explanations of professional terms and techniques, such as Figure 2 A schematic diagram of the structure of a data processing system provided as an example for this application.

[0056] The data processing system includes multiple computing clusters. Taking any computing cluster as an example, the computing cluster includes multiple computing devices (computer nodes), each computing device is connected to a connection device, that is, the connection device is used to connect any two computing devices among the multiple computing devices, and the two computing devices communicate based on an ultra-low latency communication protocol. It can be considered that each computing cluster corresponds to a respective connection device, and the connection device corresponding to the computing cluster is located within the computing cluster, or the connection device corresponding to the computing cluster is located outside the computing cluster. Among them, the ultra-low latency communication protocol is, for example, a CXL protocol or a UB protocol. Or it can be understood that any computing device located in the computing cluster can perceive the memory of other computing devices in the computing cluster to which it belongs based on the ultra-low latency communication protocol, and then address the memory of all computing devices in the computing cluster to which it belongs.

[0057] Multiple computing clusters are further networked through the RDMA protocol, and any two computing clusters are connected through a network device, wherein the network device is, for example, an RDMA network interface controller (RNIC). That is, the network device is used to connect any two computing clusters in the multiple computing clusters. Exemplarily, two computing devices located in two different computing clusters each include an RNIC, and the two RNICs can communicate based on the RDMA protocol, that is, communication between the computing clusters to which the two computing devices belong is realized.

[0058] Figure 2exemplarily shows computing cluster 1 and computing cluster 2, computing cluster 1 and computing cluster 2 are connected through a network device. Further, computing cluster 1 includes computing devices 1 to 4 and connection device 1, wherein connection device 1 is used to connect any two computing devices from computing device 1 to computing device 4; computing cluster 2 includes computing devices 5 to computing device 8 and connection device 2, wherein connection device 2 is used to connect any two computing devices from computing device 5 to computing device 8. It can be understood that Figure 2 The number of computing clusters, the number of computing devices in each computing cluster, and the positional relationship between the connecting devices and the computing clusters do not constitute a limitation on the structure of the data processing system in this application.

[0059] In the data processing system, the memories of the plurality of computing devices may form a memory pool. Figure 2 In the example, the data processing system network includes computing devices 1 to 8, and computing devices 1 to 8 include memories 1 to 8 respectively. Accordingly, memories 1 to 8 constitute a memory pool.

[0060] Further, taking any computing device as an example, the computing device includes one or more processors, such as a CPU, a GPU, a neural-network processing unit (NPU), an FPGA, etc. For any processor in the computing device, the memory pool may include one or more memory forms.

[0061] Based on the access speed from fast to slow, four memory forms are provided as examples below:

[0062] (1) Device attached memory (DAM): For CPU, for example, it is the double data rate synchronous dynamic random access memory (DDR SDRAM) connected to the CPU; for GPU, for example, it is the high bandwidth memory (HBM) in the GPU chip; for NPU, for example, it is the HBM and DDR in the NPU chip. For example, the access latency of device attached memory is less than 100ns.

[0063] (2) Device local memory (DLM): For a GPU, it is the host memory in the computing device to which the GPU belongs; for a CPU, it is the expander memory in the computing device to which the CPU belongs, etc. Exemplarily, the access latency of the device local memory is in the range of 100ns to 300nm.

[0064] (3) Small network memory (SNM): the memory within the computing cluster. Figure 2 For example, computing device 1 and computing device 2 are located in computing cluster 1, and for the CPU in computing device 1, memory 2 in computing device 2 is a small network memory of the CPU in computing device 1. Exemplarily, the access latency of SNM is in the range of 300ns to 600nm.

[0065] (4) Large network memory (LNM): This is the memory across computing clusters. Figure 2 For example, computing device 1 is located in computing cluster 1, computing device 5 is located in computing cluster 2, and for the CPU in computing device 1, memory 5 in computing device 5 is the LNM of the CPU in computing device 1. Exemplarily, the access latency of SNM is greater than 600 nm.

[0066] Combination Figure 2 , Figure 3 A schematic diagram of a memory pool of a data processing system provided as an example for the present application. Among them, computing device 1 includes processor 11 and processor 12, computing device 2 includes processor 2, and computing device 5 includes processor 5. Furthermore, computing device 1 also includes a device direct memory of processor 11 (denoted as MEM11), a device direct memory of processor 12 (denoted as MEM121), and an extended memory of processor 12 (denoted as MEM122). Computing device 2 also includes a device direct memory of processor 2 (denoted as MEM2). Computing device 5 also includes a device direct memory of processor 5 (denoted as MEM5). From the perspective of processor 11, MEM11 is a device direct memory; MEM121 and MEM122 are both device local memories; MEM2 is a small network memory; MEM5 is a large network memory. Processor 11 can access MEM2 based on an ultra-low latency communication protocol, and access MEM5 through RDMA. When processor 11 accesses MEM5 based on RDMA, it can specifically access it based on RNIC1 in computing device 1 and RNIC5 in computing device 5 (the specific implementation can be described in step 606 below).

[0067] Combination Figure 2and Figure 3 The present application provides a data processing method, which is specifically an addressing method. The addressing method can be executed by a first connection device corresponding to a first computing cluster. The first connection device can uniformly manage the memory address information of the first computing cluster and the memory address information of the second computing cluster to enable a first computing device in the first computing cluster to access a second computing device in the second computing cluster.

[0068] Exemplarily, the first computing cluster may be Figure 2 The computing cluster 1 in the example, the second computing cluster can be Figure 2 In the computing cluster 2, accordingly, the first computing device can be any one of the computing devices 1 to 4, and the second computing device can be any one of the computing devices 5 to 8.

[0069] Alternatively, the first computing cluster in the data processing method may be Figure 2 The second computing cluster can be Figure 2 In the computing cluster 1 , accordingly, the first computing device may be any one of the computing devices 5 to 8 , and the second computing device may be any one of the computing devices 1 to 4 .

[0070] Figure 4 A flowchart of a data processing method is provided for this application:

[0071] Step 401: The first computing device obtains memory address information of the third computing device.

[0072] The third computing device is a computing device other than the first computing device in the first computing cluster, combined with Figure 2 In the example, the first computing cluster is computing cluster 1, the first computing device is computing device 1, and the third computing device is computing device 2. The first computing device is connected to the third computing device via a first connecting device.

[0073] The third computing device may include multiple memories. For example, the third computing device includes a DDR SDRAM connected to the CPU, an HBM in the GPU chip, and an extended memory. The memory address information of the third computing device may specifically be information of the multiple memories, such as the size and type of each memory in the multiple memories.

[0074] Step 402: The first computing device addresses the memory address information of the first computing cluster (recorded as first memory address information) according to the memory address information of the third computing device and the memory address information of the first computing device.

[0075] The first computing device may include multiple memories. For example, the first computing device includes a DDR SDRAM connected to a CPU, an HBM in a GPU chip, and an extended memory. The memory address information of the first computing device may specifically be information of the multiple memories, such as the size and type of the multiple memories.

[0076] Specifically, the first computing device performs addressing according to the memory address information of all computing devices in the first computing cluster (including the memory address information of the third computing device and the memory address information of the first computing device) to form the memory address information of the first computing cluster from the perspective of the first computing device (i.e., the first memory address information). Figure 2 For example, the first computing cluster is computing device 1. Computing device 1 can obtain memory address information of computing devices 1 to 4, and perform addressing according to the memory address information of computing devices 1 to 4. For example, computing devices 1 to 4 all have 25G memory, that is, computing cluster 1 has a total of 100G memory, then computing device 1 can address the 100G memory from the perspective of the first computing device.

[0077] It can also be understood that, since the computing devices in the first computing cluster are connected based on the ultra-low latency communication protocol, the first computing device can sense the memory of other computing devices in the first computing cluster except itself, and then address the memory of all computing devices in the first computing cluster to obtain the first memory address information. Furthermore, for other computing devices in the first computing cluster (such as the third computing device), it is also possible to sense the memory of other computing devices in the first computing cluster except itself, and then address the memory of all computing devices in the first computing cluster to obtain the memory address information of the first computing cluster from the perspective of the other computing devices.

[0078] It should be pointed out that the two computing devices are addressed from their own perspectives, and the memory address information of the first computing cluster obtained by the two is different. Figure 2 For example, computing device 1 may consider that the address numbering of its local 25G memory is 1G to 25G, and that the address numbering of computing device 2's local 25G memory is 26G to 50G; and computing device 2 may consider that the address numbering of its local 25G memory is 1G to 25G, and that the address numbering of computing device 1's local 25G memory is 26G to 50G.

[0079] Step 403: The first computing device sends the first memory address information to the first connecting device. Correspondingly, the first connecting device receives the first memory address information from the first computing device.

[0080] In addition, other computing devices in the first computing cluster will also send the memory address information of the first computing cluster obtained by addressing to the first connection device after addressing. Accordingly, the first connection device can store the memory address information of the first computing cluster obtained by addressing from the perspective of each computing device in the first computing cluster. Figure 2 For example, the memory address information of the first computing cluster obtained by respectively addressing computing devices 1 to 4 is recorded as memory address information 1 to memory address information 4, and accordingly, the first connection device records memory address information 1 to memory address information 4. In the present application, the memory address information of the first computing cluster obtained from the perspective of each computing device in the first computing cluster and recorded in the first connection device can be referred to as the memory address information set of the first computing cluster.

[0081] Similarly, each computing device in the second computing cluster can also address the addresses of all computing devices in the second computing cluster to obtain the memory address information of the second computing cluster from the perspective of each computing device in the second computing cluster. Correspondingly, the second connection device records the memory address information of the second computing cluster from the perspective of each computing device in the second computing cluster. For specific implementation methods, please refer to the description in the above steps 401 to 403. Figure 2 For example, the memory address information of the second computing cluster obtained by respectively addressing computing devices 5 to 8 is recorded as memory address information 5 to memory address information 8, and accordingly, the second connection device records memory address information 5 to memory address information 8. In the present application, the memory address information of the second computing cluster obtained from the perspective of each computing device in the second computing cluster and recorded in the second connection device can be referred to as the memory address information set of the second computing cluster.

[0082] Furthermore, the first connection device can also obtain the memory address information set of the second computing cluster recorded in the second connection device. In one example, the first connection device is directly connected to the second connection device, and the first connection device can directly synchronize the memory address information set of the second computing cluster recorded in the second connection device from the second connection device. In another example, the first connection device is connected to the first computing device, and the second connection device is connected to the second computing device. The first computing device and the second computing device are connected through a network device (such as RNIC). Exemplarily, the second connection device first synchronizes the memory address information set of the second computing cluster to the second computing device based on an ultra-low latency communication protocol, and the second computing device then synchronizes the memory address information set of the second computing cluster to the first computing device based on the RDMA protocol. Subsequently, the first computing device synchronizes the memory address information set of the second computing cluster to the first connection device based on the ultra-low latency communication protocol. In combination with Figure 2For example, the first connection device can obtain memory address information 5 to memory address information 8 recorded in the second connection device, that is, the first connection device can manage memory address information 1 to memory address information 8.

[0083] In addition, each computing device in the second computing cluster may synchronize the memory address information in the computing device to the first connection device through the RDMA protocol. Figure 2 For example, computing device 5 synchronizes memory address information 5 to the first connected device, computing device 6 synchronizes memory address information 6 to the first connected device, computing device 7 synchronizes memory address information 7 to the first connected device, and computing device 8 synchronizes memory address information 8 to the first connected device. In this way, the first connected device can manage memory address information 1 to memory address information 8.

[0084] It should be noted that the above Figure 4 In the relevant embodiments, when the first computing device addresses the first computing cluster, specifically, the first computing device addresses the memory of the first computing cluster at the operating system level (or hardware level). Furthermore, the first connection device can also address the sum of the memories of multiple computing clusters in the entire data processing system to obtain memory address information at the application level. Subsequently, the first connection device establishes a mapping relationship between the memory address information at the application level and the memory address information of each computing device at the operating system level.

[0085] Combination Figure 2 For example, the following Table 1 is a mapping relationship for the management of a first connection device provided by the present application, wherein each address from address 1 to address n+m corresponds to the same address length, and the address length is, for example, 4kb, 8kb, etc. Computing cluster 1 corresponds to address 1 to address n at the application level, that is, the length of address 1 to address n at the application level is 100G, n is a positive integer, and n is specifically the ratio of 100G to the address length. Computing cluster 2 corresponds to address n+1 to address n+m at the application level, that is, the length of address n+1 to address n+m at the application level is 100G, m is a positive integer, and m is specifically the ratio of 100G to the address length.

[0086] Further, from the perspective of different computing devices, the same application-level address corresponds to different addresses at the operating system level. Exemplarily, address 1 at the application level corresponds to address 1-1 in computing device 1, to address 2-1 in computing device 2, to address 3-1 in computing device 3, and to address 4-1 in computing device 4. Further exemplarily, address n+1 at the application level corresponds to address 5-1 in computing device 5, to address 6-1 in computing device 6, to address 7-1 in computing device 7, and to address 8-1 in computing device 8.

[0087] Table 1

[0088]

[0089] Further, the first connection device may send multiple application-level addresses corresponding to the first computing cluster and the length of each address to the first computing device, such as sending the above-mentioned address 1 to address n, a total of n addresses, and the length of each address (such as 4kb) to the first computing device. Alternatively, the first connection device may send the starting address and total length of the application level corresponding to the first computing cluster to the first computing device, such as sending the above-mentioned address 1 and 100G to the first computing device. In addition, the first connection device may also send multiple application-level addresses of the second computing cluster and the length of each address to the first computing device, or send the starting address and total length of the application level corresponding to the second computing cluster to the first computing device. In this application, the information sent by the first connection device to the first computing device is collectively referred to as the first information. Accordingly, the first computing device can determine whether the memory space indicated by the address to be accessed by the application is located in the first computing cluster based on the first information from the first connection device (the specific implementation can be seen in the description of step 602 below).

[0090] Similarly, the second connection device can also obtain the memory address information of the first computing cluster from the perspective of each computing device in the first computing cluster recorded in the first connection device, and address the sum of the memories of multiple computing clusters in the entire data processing system to obtain memory address information at the application level. The application-level memory address information can be used for applications running on computing devices in the second computing cluster to access the entire data processing system. Furthermore, the second connection device will also send multiple application-level addresses corresponding to the second computing cluster and the length of each address to the first computing device, or send the application-level starting address and total length corresponding to the second computing cluster to the second computing device. The specific implementation is similar to that of the first connection device mentioned above and will not be repeated here.

[0091] It should be added that there can be K computing clusters in the data processing system of the present application, where K is a positive integer. For a computing cluster, the connection device corresponding to the computing cluster can record the memory address information set of the computing cluster. In order to realize that the connection device corresponding to each computing cluster can obtain the memory address information set in the connection devices of all other computing clusters, the present application provides a ring synchronization method, that is, the K connection devices corresponding to the K computing clusters can form multiple ring structures, each ring structure includes multiple connection devices that can communicate with each other based on the ultra-low latency communication protocol, so that the multiple connection devices in the ring structure can exchange their respective memory address information sets based on the ultra-low latency communication protocol. Subsequently, each of the multiple ring structures further provides a connection device, that is, a connection device provided by each of the multiple ring structures further forms a new ring structure based on the RDMA protocol, and the multiple connection devices in the new ring structure can exchange the memory address information sets of their respective ring structures based on the RDMA protocol.

[0092] like Figure 5 A schematic diagram of a ring synchronization method provided as an example for the present application, wherein the thin solid circle refers to the connection device and the thick solid circle refers to the network device. Among them, connection devices 1 to 8 form ring 1, connection devices 9 to 16 form ring 2, connection devices 17 to 24 form ring 3, connection devices 25 to 32 form ring 4, and connection devices 1, connection devices 9, connection devices 17 and connection devices 25 form ring 5. For the convenience of description, the memory address information sets of the computing clusters corresponding to connection devices 1 to 32 are respectively recorded as memory address information set 1 to memory address information set 32.

[0093] Take Ring 1 as an example to explain ring transmission:

[0094] In the first ring transfer:

[0095] Connection device 1 sends memory address information set 1 to connection device 2, connection device 2 sends memory address information set 2 to connection device 3, ..., connection device 7 sends memory address information set 7 to connection device 8, and connection device 8 sends memory address information set 8 to connection device 1.

[0096] In the second ring transfer:

[0097] Connection device 1 sends memory address information set 8 to connection device 2, connection device 2 sends memory address information set 1 to connection device 3, ..., connection device 7 sends memory address information set 6 to connection device 8, and connection device 8 sends memory address information set 7 to connection device 1.

[0098] By analogy, each connection device from connection device 1 to connection device 8 can obtain memory address information set 1 to memory address information set 8.

[0099] Similarly, in the ring transmission corresponding to ring 2, each connection device from connection device 9 to connection device 16 can obtain memory address information set 9 to memory address information set 16. In the ring transmission corresponding to ring 3, each connection device from connection device 17 to connection device 24 can obtain memory address information set 17 to memory address information set 24. In the ring transmission corresponding to ring 4, each connection device from connection device 25 to connection device 32 can obtain memory address information set 25 to memory address information set 32.

[0100] After rings 1 to 4 have completed the ring transmission, connection device 1, connection device 9, connection device 17 and connection device 25 in ring 5 further perform ring transmission (in a similar manner to ring 1), so that connection device 1, connection device 9, connection device 17 and connection device 25 can all obtain memory address information set 1 to memory address information set 32. Subsequently, connection device 1 sends memory address information set 1 to memory address information set 32 ​​to each connection device in ring 1, connection device 9 sends memory address information set 1 to memory address information set 32 ​​to each connection device in ring 2, connection device 17 sends memory address information set 1 to memory address information set 32 ​​to each connection device in ring 3, and connection device 25 sends memory address information set 1 to memory address information set 32 ​​to each connection device in ring 4.

[0101] In this way, the connection device corresponding to each computing cluster can manage not only the memory address information set of the computing cluster, but also the memory address information set of other computing clusters. When a computing device needs to access computing devices in other computing clusters other than the cluster to which it belongs, the computing device can first query the address it needs to access from the connection device corresponding to the computing cluster to which it belongs, and then access it based on the queried address.

[0102] like Figure 6 A flowchart of a data processing method exemplarily provided in the present application is provided, wherein the data processing method is specifically a data access method in which a first computing device accesses a memory space in a second computing device.

[0103] Step 601: A first computing device receives an access request, wherein the access request includes a first address to be accessed.

[0104] Step 602 , the first computing device determines whether the memory space indicated by the first address is located in the first computing cluster, and if so, executes step 603 , otherwise, executes steps 604 to 606 .

[0105] In a specific implementation, the application in the first computing device generates an access request based on the memory address information at the application level, and the access request includes a first address to be accessed by the application, and the first address is used to indicate an address space in a computing cluster of the data processing system. Figure 2 In the example, the total memory of computing cluster 1 is 100G, and the total memory of computing cluster 2 is 100G, then the total memory of the entire data processing system is 200G. The first computing device can provide 200G of memory space for the application, that is, the application can access any address space in the 200G.

[0106] In one possible implementation, the first computing device determines whether the memory space indicated by the first address is located in the first computing cluster based on the first information. In conjunction with the example in Table 1 above, the first computing device is computing device 1, and the first computing cluster is computing cluster 1. When the first address is address 1 at the application level, computing device 1 determines that the memory space indicated by the first address is located in computing cluster 1 based on the first information; when the first address is address n+1 at the application level, computing device 1 determines that the memory space indicated by the first address is not located in computing cluster 1 based on the first information. Further, the first computing device can determine that the memory space indicated by the first address is located in computing cluster 2 based on the first information, that is, computing cluster 2 is the second computing cluster.

[0107] Step 603: The first computing device accesses the memory space in the first computing cluster according to the first address.

[0108] The memory space indicated by the first address may be located in the first computing device, or in other computing devices in the first computing cluster except the first computing device. In conjunction with the example in Table 1 above, the first computing device is computing device 1, and the first computing cluster is computing cluster 1. When the first address is address 1 at the application level, computing device 1 may access address 1-1 according to address 1 at the application level, wherein the memory space indicated by address 1-1 is located in computing device 1; when the first address is address n at the application level, computing device 1 may access address 1-n according to address n at the application level, wherein the memory space indicated by address 1-n is located in computing device 4, and so on.

[0109] The first computing device accesses the memory space indicated by the first address through an ultra-low latency communication protocol. Specifically, when the access request is a write request, the access request also includes target data to be written to the first address, and the first computing device can store the target data in the memory space indicated by the first address. When the access request is a read request, the first computing device can load the target data from the memory space indicated by the first address.

[0110] Step 604: The first computing device sends a query request to the first connecting device. Correspondingly, the first connecting device receives the query request from the first computing device, wherein the query request includes the first address.

[0111] Step 605: The first connecting device sends a query response to the first computing device. Correspondingly, the first computing device receives the query response from the first connecting device, wherein the query response includes the second address.

[0112] Specifically, the first connecting device determines the second address corresponding to the first address based on the first address in the query request and the mapping relationship between the memory address information at the application level and the memory address information of each computing device at the operating system level. The second address is used for the first computing device to access the second computing device.

[0113] In conjunction with the example in Table 1 above, when the first address is the address n+1 at the application level, the computing device 1 sends a query request to the first connection device. The first connection device determines that the second address is specifically address 5-1 from the perspective of computing device 5 based on the address n+1 at the application level in the query request and the mapping relationship in Table 1, and then sends a query response to the computing device 1, the query response including the identifier of computing device 5 and address 5-1; or, the first connection device can also determine that the second address is specifically address 6-1 from the perspective of computing device 6, and then sends a query response to the computing device 1, the query response including the identifier of computing device 6 and address 6-1. It can be understood that the first connection device can determine the second address corresponding to the first address from the perspective of multiple different computing devices based on the first address. Further, the first connection device can select a computing device with a lower computing load, and send the identifier of the computing device with a lower computing load and the second address corresponding to the computing device to the first computing device.

[0114] Step 606: The first computing device accesses the memory space in the second computing device according to the second address.

[0115] Specifically, the first computing device accesses the second address in the second computing device based on the RDMA protocol.

[0116] When the access is a write request (i.e., RDMA write), the access request also includes the target data to be written to the first address. The CPU of the second computing device first registers the second address in the RNIC of the second computing device in advance, and obtains the local key, where the local key is used to indicate that RDMA has the authority to operate the second address. Furthermore, the RNIC of the second computing device encapsulates the second address and the local key into a special message and transmits it to the first computing device, that is, the RNIC of the second computing device transfers the operation authority of the second address to the first computing device. Correspondingly, after the RNIC of the first computing device receives the special message sent from the RNIC of the second computing device, it encapsulates the second address, the local key, and the target data stored in the first computing device into the RDMA write and sends it to the second computing device. In this way, the first computing device can write the target data to the second address of the second computing device.

[0117] When the access is a read request (i.e., RDMAread), the CPU of the second computing device first registers the second address in advance to the RNIC of the second computing device, and obtains the local key, where the local key is used to indicate that RDMA has the authority to operate the second address. Furthermore, the RNIC of the second computing device encapsulates the second address and the local key into a special message and transmits it to the first computing device, that is, the second computing device transfers the operation authority of the second address to the first computing device. Correspondingly, after the RNIC of the first computing device receives the special message sent from the RNIC of the second computing device, it encapsulates the third address, the second address, and the local key used to store the target data locally into RDMA read and sends it to the second computing device. In this way, the first computing device can read the target data from the second address of the second computing device to the third address of the first computing device.

[0118] It should be added that the first computing device can also monitor the occupancy of each memory locally. For example, the first computing device includes a DDR SDRAM linked to the CPU, an HBM in the GPU chip, and an extended memory. The first computing device monitors that the occupancy rate of the DDR SDRAM is 70%, the occupancy rate of the HBM is 50%, and the occupancy rate of the extended memory is 20%. Similarly, other computing devices can also monitor the occupancy of each local memory. Subsequently, the first computing device can obtain the occupancy of each memory of each computing device in the computing cluster to which it belongs, as well as the occupancy of each memory of each computing device in other computing clusters. When the first computing device obtains an access request (specifically a write request) from an application, it can determine which memory to write the target data in the write request to based on the occupancy of each memory in the entire data processing system. Exemplarily, the first computing device comprehensively considers the occupancy of each memory, as well as the latency of the first computing device accessing each memory (i.e., the latency of accessing the device's directly connected memory, the latency of accessing the device's local memory, the latency of accessing the small network memory, and the latency of accessing the large network memory), and selects a memory from multiple memories into which to write the target data. When the selected memory is located in the first computing cluster, the first computing device can store the target data in the memory based on an ultra-low latency communication protocol; when the selected memory is located in another computing cluster (such as the second computing cluster), the first computing device can write the target data to the memory based on the RDMA protocol. Furthermore, the first computing device obtains the memory occupancy of each memory of other computing devices in the computing cluster to which it belongs, and the implementation method for the first computing device to obtain the memory occupancy of each computing device in other computing clusters can be found in the above. Figure 5 As described in the relevant embodiments, at this time, the thin solid line circle refers to the computing device, and the thick solid line circle refers to the network device.

[0119] It should also be added that the first computing device can also monitor the frequency at which the data stored in each local memory is read by the first computing device. Similarly, other computing devices can also monitor the frequency at which the data stored in each local memory is read by the first computing device. Subsequently, the first computing device can obtain the frequency at which the data stored in each memory of other computing devices in the computing cluster to which it belongs is read by the first computing device, and obtain the frequency at which the data stored in each memory of each computing device in other computing clusters is read by the first computing device. When the first computing device determines that the frequency at which a certain data is read by the first computing device is inconsistent with the frequency interval corresponding to the memory in which it is located, the first computing device can move the data to the memory corresponding to the frequency interval to which the frequency of the data being read belongs. It can be understood that each memory corresponds to a different frequency interval, for example, the DDR SDRAM of computing device 1 corresponds to frequency interval 1, the HBM of computing device 1 corresponds to frequency interval 2, the extended memory in computing device 1 corresponds to frequency interval 3, the memory in computing device 2 corresponds to frequency interval 4, and the memory in computing device 5 corresponds to frequency interval 5, wherein frequency interval 1, frequency interval 2, frequency interval 3, frequency interval 4, and frequency interval 5 are successively reduced. For example, the first computing device detects that data 1 is located in the memory of computing device 5, and the frequency at which the first computing device reads data 1 is in frequency interval 1, then the first computing device needs to move data 1 to the DDR SDRAM of computing device 1; for another example, the first computing device detects that data 2 is located in the HBM of computing device 1, and the frequency at which the first computing device reads data 2 is in frequency interval 4, then the first computing device needs to move data 2 to the memory of computing device 2. In this way, the first computing device can place data with a higher reading frequency in a memory with a lower reading latency, which helps to improve the efficiency of data reading. Furthermore, the first computing device obtains the frequency at which the data stored in each memory of other computing devices in the computing cluster to which it belongs is read by the first computing device, and the first computing device obtains the frequency at which the data stored in each memory of each computing device in other computing clusters is read by the first computing device, which can be seen in the above. Figure 5 As described in the relevant embodiments, at this time, the thin solid line circle refers to the computing device, and the thick solid line circle refers to the network device.

[0120] Based on the same inventive concept as above, the present application provides a possible data processing system, which includes: a first computing cluster, including a first computing device; a second computing cluster, including a second computing device; a first connecting device, connected to the first computing device, for managing memory address information of the first computing cluster provided by the first computing device, and memory address information of the second computing cluster provided by the second computing device; the first computing device is used to: receive an access request, the access request is used to access the memory space of the second computing device; search for the address of the memory space from the memory address information of the second computing cluster managed by the first connecting device; and access the memory space in the second computing device according to the address.

[0121] In a possible implementation, the data processing system further includes: a network device configured to connect the first computing cluster and the second computing cluster.

[0122] In a possible implementation manner, the first connection device is connected to the first computing device through the CXL protocol or the UB protocol.

[0123] In one possible implementation, the first computing cluster also includes a third computing device, and the first connecting device is used to connect the first computing device and the third computing device; the first computing device is also used to: obtain memory address information of the third computing device; address the memory address information of the first computing cluster according to the memory address information of the third computing device and the memory address information of the first computing device; and send the memory address information of the first computing cluster to the first connecting device.

[0124] In a possible implementation, the memory address information of the first computing device includes information of multiple memories in the first computing device; and the memory address information of the third computing device includes information of multiple memories in the third computing device.

[0125] In a possible implementation, the first connecting device is further used to: receive memory address information of a first computing cluster from the first computing device, and / or receive memory address information of a second computing cluster from the second computing device.

[0126] Based on the same inventive concept as above, Figure 7 The data processing device 70 provided in the embodiment of the present application is shown. The data processing device 70 can be applied to the flowchart shown above to perform the functions of the first computing device or the first connecting device in the above method embodiment. For the convenience of explanation, Figure 7 Only the main components of the device are shown.

[0127] The data processing device 70 includes a processor 701 , a memory 702 and a communication interface 703 . Any two of the processor 701 , the memory 702 and the communication interface 703 may be connected via a bus 704 .

[0128] The processor 701 may be a central processing unit (CPU), which may be used to execute program instructions in the memory 702 to implement the above Figures 4 to 6 Operations in the related method embodiments. In addition to the CPU, the processor 701 may also be an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), a system on chip (SoC) or a complex programmable logic device (CPLD), a graphics processing unit (GPU), a neural-network processing unit (NPU), etc.

[0129] It should be noted that in actual applications, there may be multiple processors 701, and the multiple processors 701 may include multiple processors of the same type, or may include multiple processors of different types. For example, multiple processors 701 are multiple CPUs. For another example, the multiple processors 701 include one or more CPUs and one or more GPUs. For another example, the multiple processors 701 include one or more CPUs and one or more NPUs. Alternatively, the multiple processors 701 include one or more CPUs, one or more GPUs, and one or more NPUs, etc. Among them, the processor 701 (such as a CPU, an NPU, etc.) may include one core, or may include multiple cores.

[0130] The memory 702 refers to a device for storing program instructions and / or data, which can be a memory or a hard disk.

[0131] Memory refers to an internal memory that directly exchanges data with the processor 701. It can read and write data at any time and at a very fast speed. It serves as a temporary data storage for the operating system or other running programs running on the processor 701. Memory includes volatile memory (volatile memory), for example, random access memory (RAM), dynamic random access memory (DRAM), etc., and may also include non-volatile memory (non-volatile memory), such as storage class memory (SCM), etc., or a combination of volatile memory and non-volatile memory. In practical applications, multiple memories can be configured in the data processing device 70, and optionally, the multiple memories can be of different types. This embodiment does not limit the number and type of memories. In addition, the memory can be configured to have a power-saving function. The power-saving function means that when the system loses power and then powers on again, the data stored in the memory will not be lost. Memory with a power-saving function is called a non-volatile memory.

[0132] Hard disk, including but not limited to: non-volatile memory (non-volatile memory), such as read-only memory (ROM), hard disk drive (HDD) or solid state drive (SSD), etc. Unlike memory, hard disk has a slower read and write speed and is usually used to store data persistently. In one embodiment, the data, program instructions, etc. in the hard disk need to be loaded into the memory first, and then the processor obtains these data and / or program instructions from the memory.

[0133] The communication interface 703 is used to communicate with other devices. For example, when the data processing device 70 is a first computing device, the communication interface 703 is used by the first computing device to receive access requests, send query requests to the connection device, receive query responses from the connection device, etc.; when the data processing device 70 is a connection device, the communication interface 703 is used by the connection device to receive query requests from the first computing device, and send query responses to the first computing device, etc.

[0134] Based on the same inventive concept as above, the present application provides a computer-readable storage medium, in which a computer program or instruction is stored. When the computer program or instruction is executed by a device, the above Figures 4 to 6 The method in the related method embodiment.

[0135] Based on the same inventive concept as above, the present application provides a computer program product, which includes a computer program or instructions. When the computer program or instructions are executed by a device, the above Figures 4 to 6 The method in the related method embodiment.

[0136] Based on the same inventive concept as above, the present application provides a device, including a processor, the processor is connected to a memory, the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory, so that the device can achieve the above Figures 4 to 6 The method in the related method embodiment.

[0137] Based on the same inventive concept as above, the present application provides a data processing device, which can be specifically Figures 4 to 6 The first computing device or the first connecting device in the related method embodiment, correspondingly, when the data processing device is the first computing device, the data processing device may include a module for executing the above-mentioned first computing device function (such as a transceiver module, a search module and an access module); when the data processing device is the first connecting device, the data processing device may include a module for executing the above-mentioned first connecting device function (such as a transceiver module and a management module).

[0138] In the present application, "at least one" means one or more, and "plurality" means two or more. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single items or plural items. For example, at least one of a, b or c can be represented by: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, c can be single or multiple. "And / or" describes the association relationship of associated objects, indicating that three relationships can exist. For example, A and / or B can be represented by: A exists alone, A and B exist at the same time, and B exists alone, where A and B can be singular or plural. In the text description of the present application, the character " / " generally indicates that the previous and next associated objects are in an "or" relationship; in the formula of the present application, the character " / " indicates that the previous and next associated objects are in a "division" relationship.

[0139] It is understood that the various numbers involved in the embodiments of the present application are only for the convenience of description and are not used to limit the scope of the embodiments of the present application. The size of the sequence number of the above-mentioned processes does not mean the order of execution, and the execution order of each process should be determined by its function and internal logic.

[0140] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the scope of protection of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A data processing system, It is characterized in that include: A first computing cluster, comprising a first computing device; a second computing cluster, comprising a second computing device; a connection device, connected to the first computing device and the second computing device, and used to manage memory address information of computing devices in the second computing cluster; The first computing device is used to: receiving an access request, the access request being used to access a memory space of the second computing device; Find the address of the memory space from the memory address information of the computing device in the second computing cluster managed by the connection device; The memory space in the second computing device is accessed according to the address.

2. The system according to claim 1, It is characterized in that Also includes: A network device is used to connect the first computing cluster and the second computing cluster.

3. The system according to claim 1, It is characterized in that The connection device is connected to the first computing device via an open interconnect CXL protocol or a unified bus UB protocol.

4. The system according to claim 1, It is characterized in that The first computing cluster further includes a third computing device, and the connection device is used to connect the first computing device and the third computing device; The first computing device is further configured to: Obtaining memory address information of the third computing device; Addressing the memory address information of the first computing cluster according to the memory address information of the third computing device and the memory address information of the first computing device; The memory address information of the first computing cluster is sent to the connecting device.

5. The system according to claim 4, It is characterized in that The memory address information of the first computing device includes information of a plurality of memories in the first computing device; The memory address information of the third computing device includes information of multiple memories in the third computing device.

6. The system of claim 1, It is characterized in that The first computing device is further configured to: After receiving the access request, based on the fact that the address accessed by the access request does not belong to the memory address of the first computing cluster, it is determined that the access request is an access request for accessing the memory space of the second computing device.

7. The system according to any one of claims 1 to 6, It is characterized in that The connection device is also used for: receiving memory address information of the first computing cluster from the first computing device, and / or, Receive memory address information of the second computing cluster from the second computing device.

8. A data processing method, It is characterized in that include: A first computing device in a first computing cluster receives an access request, where the access request is used to access a memory space of a second computing device in a second computing cluster; Sending a query request to a connection device, the query request being used to query the address of the memory space of the second computing device, wherein the connection device includes memory address information of computing devices in the second computing cluster; The connecting device receives the query request, searches for the address of the memory space of the second computing device according to the memory address information of the computing device in the second computing cluster according to the query request, and sends the address obtained by the query to the first computing device; The first computing device accesses the memory space in the second computing device according to the received address.

9. The method according to claim 8, It is characterized in that Also includes: The connection device manages memory address information of the first computing cluster provided by the first computing device and memory address information of the second computing cluster provided by the second computing device.

10. The method according to claim 8, It is characterized in that The connection device is connected to the first computing device via an open interconnect CXL protocol or a unified bus UB protocol.

11. The method according to claim 8, It is characterized in that The first computing cluster further includes a third computing device, and the connection device is used to connect the first computing device and the third computing device; The method further comprises: The first computing device obtains memory address information of the third computing device; The first computing device addresses the memory address information of the first computing cluster according to the memory address information of the third computing device and the memory address information of the first computing device; The first computing device sends the memory address information of the first computing cluster to the connecting device.

12. The method according to claim 11, It is characterized in that The memory address information of the first computing device includes information of a plurality of memories in the first computing device; The memory address information of the third computing device includes information of multiple memories in the third computing device.

13. The method according to claim 8, It is characterized in that After the first computing device receives the access request, the method further includes: The first computing device determines that the access request is an access request for accessing the memory space of the second computing device based on the fact that the address accessed by the access request does not belong to the memory address of the first computing cluster.

14. The method according to any one of claims 8 to 13, It is characterized in that Also includes: The connecting device receives the memory address information of the first computing cluster from the first computing device, and / or receives the memory address information of the second computing cluster from the second computing device.

15. A connection device, It is characterized in that include: A transceiver module, configured to receive a query request sent by a first computing device in a first computing cluster, wherein the query request is used to query an address of an address space of a second computing device in a second computing cluster, wherein the connection device includes memory address information of the computing device in the second computing cluster; The management module is used to search the address of the memory space of the second computing device in the memory address information of the computing device in the second computing cluster according to the query request, and send the address obtained by the query to the first computing device.

16. The connection device according to claim 15, It is characterized in that The transceiver module is also used for: receiving memory address information of the first computing cluster from the first computing device, and / or, Receive memory address information of the second computing cluster from the second computing device.

Citation Information

Patent Citations

  • Data processing method and device based on cluster access

    CN111026709A

  • Data processing method and device based on distributed shared memory system

    CN113590364A

  • Data transmission method, processor system and memory access system

    CN113852656A

  • Data transmission method and device for computing cluster and storage medium

    CN114153754A

  • Storage cluster interconnection method and device, computer equipment and storage medium

    CN114357049A