Database query task RDMA unloading method and system based on multi-stage assembly line

Abstract

The invention provides a database query task RDMA unloading method and system based on a multistage assembly line, and relates to the technical field of RDMA networks and database query acceleration. Forwarding and unloading of database query task data are directly completed through an RDMA network and memory units of different computing node units; wherein the calculation of the database query task comprises a single-task operation mode and a multi-task fusion operation mode, and when the issued task is in the multi-task fusion operation mode, corresponding calculation node units are selected according to basic operations contained in multiple tasks to form pipeline operation; after the calculation result generated in the middle is directly forwarded to the memory unit corresponding to the calculation node unit, the calculation result is directly cached to the memory unit of the next-level calculation node unit through the RDMA network, the next-level calculation node unit reads the middle result in the memory unit to continue operation, and so on until the final result is returned to the CPU memory of the host node.

Description

Technical Field

[0001] The present disclosure relates to the technical field of RDMA network and database query acceleration, and in particular to a method and system for RDMA offloading of database query tasks based on a multi-stage pipeline. Background Art

[0002] The statements in this section merely provide background information related to the present disclosure and do not necessarily constitute prior art.

[0003] In recent years, database systems have become comprehensive application platforms that integrate three basic technologies: computing, storage, and networking. Currently, database systems are gradually moving towards offloading data-intensive and computing-intensive operations to dedicated hardware platforms to achieve more efficient data computing.

[0004] Traditional single heterogeneous computing nodes have obvious computing performance improvements for single operations in database query operations. However, when a query task contains multiple basic operations and needs to be executed in a pipeline between computing nodes, the computing acceleration performance is significantly reduced due to the frequent interaction of data between the CPU memory and the computing node memory. That is, the computing nodes are independent of each other and when data interaction occurs, it must first be read back to the CPU memory and then transmitted to the corresponding host node CPU memory through the traditional IP / UDP network. Direct data forwarding cannot be achieved, resulting in frequent data replication between the CPU memory and the computing node memory. Summary of the invention

[0005] In order to solve the above problems, the present disclosure proposes a database query task RDMA offloading method and system based on a multi-stage pipeline. By forming a distributed computing system with computing units, memory units and RDMA network units in different computing nodes, data can be directly forwarded between computing nodes, avoiding frequent data replication between the memory of the host node CPU and the computing nodes during pipeline computing, thereby greatly improving the computing efficiency of database query operations.

[0006] According to some embodiments, the present disclosure adopts the following technical solutions: The RDMA offloading method for database query tasks based on multi-stage pipeline includes: Build a distributed computing node architecture centered on the RDMA network unit, including computing units, memory units, and RDMA network units; Based on the distributed computing node architecture centered on the RDMA network unit, the forwarding and unloading of database query task data is directly completed through the RDMA network unit and the memory units of different computing units without CPU intervention; Among them, the calculation of database query tasks includes single-task operation mode and multi-task fusion operation mode. When the task issued is the multi-task fusion operation mode, the corresponding computing unit is selected according to the basic operations contained in the multi-task to form a pipeline operation. The intermediate calculation results are directly forwarded to the memory unit corresponding to the computing unit, and then directly cached to the memory unit of the next-level computing unit via the RDMA network. The next-level computing unit reads the intermediate results in the memory unit to continue the calculation, and so on until the final result is returned to the CPU memory of the host node.

[0007] According to some embodiments, the present disclosure adopts the following technical solutions: The RDMA offload system for database query tasks based on multi-stage pipeline includes: Computing node architecture building module, used to build a distributed computing node architecture centered on RDMA network units; The database query task offloading module is used to directly complete the forwarding offloading of database query task data through the RDMA network and the memory units of different computing units without CPU intervention based on the distributed computing node architecture centered on the RDMA network unit; Among them, the calculation of database query tasks includes single-task operation mode and multi-task fusion operation mode. When the task issued is the multi-task fusion operation mode, the corresponding computing unit is selected according to the basic operations contained in the multi-task to form a pipeline operation. The intermediate calculation results are directly forwarded to the memory unit corresponding to the computing unit, and then directly cached to the memory unit of the next-level computing unit via the RDMA network. The next-level computing unit reads the intermediate results in the memory unit to continue the calculation, and so on until the final result is returned to the CPU memory of the host node.

[0008] According to some embodiments, the present disclosure adopts the following technical solutions: A computer program product comprises a computer program, wherein when the computer program is executed by a processor, the multi-stage pipeline-based database query task RDMA offloading method is implemented.

[0009] According to some embodiments, the present disclosure adopts the following technical solutions: A non-transitory computer-readable storage medium is used to store computer instructions. When the computer instructions are executed by a processor, the multi-stage pipeline-based database query task RDMA offloading method is implemented.

[0010] According to some embodiments, the present disclosure adopts the following technical solutions: An electronic device comprises: a processor, a memory and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory, so that the electronic device executes the RDMA offloading method for database query tasks based on a multi-stage pipeline.

[0011] Compared with the prior art, the present invention has the following beneficial effects: The disclosed database query task RDMA offloading method based on multi-stage pipeline aims to solve the problem that when a database query operation includes multiple basic operations, different computing nodes frequently exchange data through the memory of the CPU, resulting in a decrease in query operation performance. The database query task RDMA offloading method based on multi-stage pipeline is proposed. The computing units, memory units and RDMA network units in different computing nodes are combined into a distributed computing system to realize direct data forwarding between computing nodes, avoid frequent data replication between the memory of the host node CPU and the computing nodes during the pipeline operation, and greatly improve the computing efficiency of the database query operation.

[0012] The disclosed RDMA offloading method for database query tasks based on multi-stage pipelines is based on RDMA network technology to form a distributed computing system architecture with computing units. The network technology allows data to be directly copied from the memory of one node to the memory of another node to achieve memory expansion and high-speed data forwarding without CPU intervention. Compared with traditional computing nodes that are independent of each other and must first be read back to the CPU memory and then transmitted to the corresponding host node CPU memory through the traditional IP / UDP network when data interaction occurs, the disclosed method allows the source computing node to directly access the memory unit of the destination computing node through the RDMA network unit in the node, realize direct data forwarding, and avoid frequent data copying between the CPU memory and the computing node memory.

[0013] The disclosed RDMA offloading method for database query tasks based on multi-stage pipelines uses a CPU+FPGA heterogeneous platform as a computing node, and uses an RDMA network to form a distributed computing system in different computing nodes, so that different computing units can directly forward data quickly and efficiently through memory units via RDMA network units, avoiding the problem of frequent data exchange between host nodes and computing nodes leading to a decrease in overall computing performance, and greatly improving the computing efficiency of database query operations. At the same time, based on the high bandwidth and low latency characteristics of the RDMA technology, the memory of the destination computing node and the memory of the source computing node form a memory pool to achieve memory expansion. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The accompanying drawings constituting a part of the present disclosure are used to provide a further understanding of the present disclosure. The illustrative embodiments of the present disclosure and their descriptions are used to explain the present disclosure and do not constitute an improper limitation on the present disclosure.

[0015] Figure 1 A schematic diagram of the composition of a computing node unit according to an embodiment of the present disclosure; Figure 2 A distributed computing node architecture centered on an RDMA network according to an embodiment of the present disclosure; Figure 3 This is a schematic diagram of multi-query task data forwarding and unloading according to an embodiment of the present disclosure; DETAILED DESCRIPTION The present disclosure is further described below in conjunction with the accompanying drawings and embodiments.

[0016] It should be noted that the following detailed descriptions are all illustrative and are intended to provide further explanation of the present disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present disclosure belongs.

[0017] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present disclosure. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0018] Example 1 In one embodiment of the present disclosure, a database query task RDMA offloading method based on a multi-stage pipeline is provided. Based on RDMA network transmission technology, a CPU+FPGA heterogeneous platform is used as a computing node, and different computing nodes are combined into a distributed computing system through an RDMA network, so that different computing node units can directly forward data quickly and efficiently through a memory unit and an RDMA network unit, including: Step 1: Build a distributed computing node architecture centered on the RDMA network unit; Step 2: Based on the distributed computing node architecture centered on the RDMA network unit, the database query task data is forwarded and unloaded directly through the RDMA network unit and the memory units of different computing units without CPU intervention; Among them, the calculation of database query tasks includes single-task operation mode and multi-task fusion operation mode. When the task issued is the multi-task fusion operation mode, the corresponding computing unit is selected according to the basic operations contained in the multi-task to form a pipeline operation. The intermediate calculation results are directly forwarded to the memory unit corresponding to the computing unit, and then directly cached to the memory unit of the next-level computing unit via the RDMA network unit. The next-level computing unit reads the intermediate results in the memory unit to continue the calculation, and so on until the final result is returned to the CPU memory of the host node.

[0019] As an embodiment, the RDMA offloading method for database query tasks based on a multi-stage pipeline disclosed in the present invention allows data to be directly copied from the memory of one node to the memory of another node to achieve memory expansion and high-speed data forwarding without CPU intervention, and allows the source computing node to directly access the memory unit of the destination computing node through the RDMA network unit in the node to achieve direct data forwarding. The specific implementation process is as follows: Step 1: Build a distributed computing node architecture centered on the RDMA network unit; Specifically, the CPU+FPGA heterogeneous platform is used as the computing node, and different computing nodes are organized into a distributed computing system through the RDMA network unit, so that different computing units can directly forward data quickly and efficiently through the memory unit via the RDMA network unit.

[0020] First, if Figure 1 As shown, the computing node architecture includes a host node, a computing unit, a memory unit, an RDMA network unit and a next-level RDMA network unit, wherein the host node Host generates a database query task, and by calling the kernel API function, transmits the relevant task type, the destination computing node and other parameters to the computing node through the PCIE bus, and at the same time, the CPU calls the XRT API to cache the data to be processed to the memory unit M via the PCIE DMA, and at the same time, deploys the corresponding computing units in different computing nodes according to the task type, and generates a start signal.

[0021] The computing node includes a computing unit, a memory unit and an RDMA network unit. The computing unit and the memory unit can send and receive data, the memory unit and the RDMA network unit can send and receive communication data, and the RDMA network unit of the computing node can send and receive data with the next-level RDMA network unit of the next-level computing unit.

[0022] Secondly, the basic operations of the database query task include sorting operation P, aggregation operation J, filtering operation G and connection operation L. The computing units of the basic operations include sorting computing unit, aggregation computing unit, filtering computing unit and connection computing unit. The computing system supports single-task operation mode and multi-task fusion operation mode. The basic operation of the multi-task fusion operation mode is the pipeline operation composed of corresponding computing node units.

[0023] Furthermore, the single-task operation mode is a single sorting operation P, aggregation operation J, filtering operation G or connection operation L, and the multi-task fusion operation mode includes: (1) Sorting operation P + aggregation operation J, (2) Sorting operation P + filtering operation G, (3) Sorting operation P + connection operation L, (4) Aggregation operation J + filtering operation G (5) Aggregation operation J + join operation L (6) Filter operation G + join operation L (7) Sorting operation P + aggregation operation J + filtering operation G, (8) Sorting operation P + aggregation operation J + join operation L, (9) Sorting operation P + filtering operation G + joining operation L, (10) Aggregation operation J + filtering operation G + connection operation L, (11) Sorting operation P + aggregation operation J + filtering operation G + joining operation L.

[0024] Without considering the execution order or dependencies, there are a total of C(4, 2)+C(4, 3)+C(4, 4)=6+4+1=11 basic combinations of multi-task fusion operations.

[0025] Step 2: Based on the distributed computing node architecture centered on the RDMA network, the database query task data is forwarded and unloaded directly through the RDMA network and the memory units of different computing node units without CPU intervention; Specifically, when the database issues a basic operation task of a single task in a query operation, the computing system schedules the corresponding computing unit to complete the operation and returns the result to the corresponding memory unit and finally returns it to the CPU memory of the host node.

[0026] When a multi-task operation fusion task is issued, the computing system needs to select the corresponding computing unit to form a pipeline operation according to the basic operations contained in the task. The intermediate calculation results are not returned to the host node memory, but forwarded to the memory unit corresponding to the computing unit, and directly cached to the memory unit of the next level computing unit via the RDMA network unit. The next level computing unit reads the intermediate results in the memory and continues to calculate, and so on until the final result is returned to the host memory. The distributed computing system composed of the RDMA network can effectively reduce the number of data copies between the CPU memory and the computing node memory, avoiding the problem of frequent data exchange between the host node and the computing node resulting in a decrease in overall computing performance, and greatly improving the computing efficiency of database query operations.

[0027] As an embodiment, the basic operation task calculation of the query operation sent by the database is completed by four computing nodes, and the specific process is as follows: S1: The host node Host generates a database query task and transfers the relevant task type, target computing node and other parameters to the computing node through the PCIE bus by calling the kernel API function. At the same time, the CPU calls the XRT API to cache the data to be processed to the memory unit M via the PCIE DMA. At the same time, the corresponding computing units are deployed in different computing nodes according to the task type and a start signal is generated.

[0028] S2: When the database issues a basic operation task of a single task in a query operation, the computing system schedules the corresponding computing unit to complete the operation and returns the result to the corresponding memory unit and finally returns it to the CPU memory of the host node.

[0029] Among them, if the task only includes a single connection operation L, the computing node deploys a connection computing unit to read the data to be calculated from the memory unit M4. The memory unit M4 is divided into multiple storage areas, namely M4-1, M4-2, M4-3, and M4-4. Among them, M4-1 and M4-2 store the inner table data and the outer table data respectively. According to the connection algorithm, the connection computing unit first reads the inner table data and stores the inner table data in M4-3 according to the hash address after hash address mapping. Then the connection computing unit reads the outer table data element in M4-2. Assuming that the outer table data element is To, the element generates a hash address H after hash address mapping. According to the address H, the inner table element Ti under the address is read from the storage area M4-3. Then, if the values ​​of To and Ti are equal, the match is successful. The merged element {To, Ti} is stored as the connection result in M4-4. This is repeated until the outer table elements are traversed. The final connection result is stored in M4-4 and an end signal is generated at the same time. After the CPU detects the end signal of the computing unit, it reads the connection result from M4-4.

[0030] S3: Assuming that the task is a multi-task operation fusion task, the computing system needs to select the corresponding computing units to form a pipeline operation according to the basic operations contained in the task. The intermediate calculation results are not returned to the host node memory, but forwarded to the memory unit corresponding to the computing unit, and directly cached to the memory unit of the next-level computing unit through the RDMA network unit. The next-level computing unit reads the intermediate results in the memory and continues the calculation, and so on until the final result is returned to the host memory.

[0031] Specifically, if two operations, filtering operation G+sorting operation P, are included, the computing node where the filtering computing unit is deployed is the master node, and the computing node where the sorting computing unit is deployed is the slave node to start the operation. The filtering computing unit acts as the master node. First, the host node caches the data to be calculated in the memory unit M2. Then, based on the filtering architecture of the truth table, the filtering computing unit compares the received column data with the corresponding column filtering conditions. The comparison results obtained are the characteristics of the row data. These results represent whether each filter column of the row meets the filtering conditions. Then, the comparison result is used as the read address of the truth table RAM. The read value of the RAM represents whether the current row data meets the filtering conditions. If the filtering conditions are met, the row data is output and cached in M2. The filtering operation of the data to be calculated is completed by analogy. According to the task type, the filtering results in M2 are transmitted to the memory unit M1 of the slave node via the RDMA network unit. Then, the sorting calculation unit is started from the node. The eight sorting subunits in it use the eight-channel merge sort algorithm to divide the data to be sorted into eight parts. The eight-channel merge sort algorithm uses multi-level comparison logic. Each level of comparison logic consists of several pairs of comparators. Each pair of comparators consists of two RAMs and one comparator. The RAM caches input data and blocks the control. The comparator compares the output data of the two RAMs. The comparator uses the empty flag signal of the previous RAM and the full flag signal of the next RAM as a valid signal to prevent the overflow of input or output data. The intermediate sorting results generated by the sorting subunit are then cached in different areas of M1. The merge sorting subunit reads the intermediate results and merges them using merge sorting to complete the final sorting and cache the results in M1. At this point, the multi-task fusion computing operation has been completed, and the RDMA network unit returns the final computing result to the memory unit M2 through the link E1. The master node reads the computing result data back to the CPU memory through the PCIE bus. After that, each node releases the resources occupied by the computing unit, removes the RDMA network link E1, and waits for the next computing task to be unloaded.

[0032] As attached Figure 2As shown in the figure, for more complex multi-task fusion operations, if it includes sorting operation P+aggregation operation J+filtering operation G+connection operation L, assuming that there are two pipelines of "sort-filter-connection" and "sort-aggregation-connection", the computing node of the sorting computing unit is taken as the master node, and the computing nodes of the sorting computing unit and the aggregation computing unit are deployed as slave nodes to start the operation, and the pipeline tasks are unloaded according to the two paths M1-E1-M2-E2-M4 and M1-E4-M3-E3-M4 respectively. Finally, the RDMA network unit returns the final operation result to the memory unit M1 through the link E5, and the master node reads the operation result data back to the CPU memory through the PCIE bus.

[0033] As attached Figure 3 As shown, the query operation of single-task computing only involves a single node, such as from CPU memory to memory unit to computing unit (such as attached Figure 3 The query operation in multi-task fusion operation is designed with two or more nodes, including multi-stage query operations (such as Figure 3 Query operation 2 and query operation 3 in ).

[0034] Example 2 In one embodiment of the present disclosure, a database query task RDMA offloading system based on a multi-stage pipeline is provided, comprising: Computing node architecture building module, used to build a distributed computing node architecture centered on RDMA network; The database query task offloading module is used to directly complete the forwarding offloading of database query task data based on the distributed computing node architecture centered on the RDMA network through the RDMA network and the memory units of different computing node units without CPU intervention; The calculation of database query tasks includes single-task operation mode and multi-task fusion operation mode. When the task issued is a multi-task fusion operation mode, the corresponding computing node unit is selected according to the basic operations contained in the multi-task to form a pipeline operation. The intermediate calculation results are directly forwarded to the memory unit corresponding to the computing node unit, and then directly cached to the memory unit of the next-level computing node unit via the RDMA network. The next-level computing node unit reads the intermediate results in the memory unit to continue the operation, and so on until the final result is returned to the CPU memory of the host node.

[0035] Example 3 In one embodiment of the present disclosure, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the multi-stage pipeline-based database query task RDMA offloading method is implemented.

[0036] Example 4 In one embodiment of the present disclosure, a non-transitory computer-readable storage medium is provided, wherein the non-transitory computer-readable storage medium is used to store computer instructions. When the computer instructions are executed by a processor, the multi-stage pipeline-based database query task RDMA offloading method is implemented.

[0037] Example 5 In one embodiment of the present disclosure, an electronic device is provided, comprising: a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory, so that the electronic device executes the RDMA offloading method for database query tasks based on a multi-stage pipeline.

[0038] The present disclosure is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present disclosure. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of the processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0039] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0040] Although the above describes the specific implementation methods of the present disclosure in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present disclosure. Technical personnel in the relevant field should understand that on the basis of the technical solution of the present disclosure, various modifications or variations that can be made by those skilled in the art without creative work are still within the scope of protection of the present disclosure.

Claims

1. A database query task RDMA offloading method based on a multi-stage pipeline, characterized in that: include: Build a distributed computing node architecture centered on the RDMA network unit, including computing units, memory units, and RDMA network units; Based on the distributed computing node architecture centered on the RDMA network unit, the forwarding and unloading of database query task data is directly completed through the RDMA network unit and the memory units of different computing units without CPU intervention; The calculation of database query tasks includes single-task operation mode and multi-task fusion operation mode. When the task issued is in multi-task fusion operation mode, the corresponding computing unit is selected according to the basic operations contained in the multi-task to form a pipeline operation. The intermediate calculation results are directly forwarded to the memory unit corresponding to the computing unit, and then directly cached to the memory unit of the next-level computing unit via the RDMA network unit. The next-level computing unit reads the intermediate results in the memory unit to continue the calculation, and so on until the final result is returned to the CPU memory of the host node.

2. The RDMA offloading method for database query tasks based on multi-stage pipeline according to claim 1, characterized in that: The basic operations of the database query task include sorting operation P, aggregation operation J, filtering operation G and connection operation L. The calculation units of the basic operations include sorting calculation unit, aggregation calculation unit, filtering calculation unit and connection calculation unit. The basic operation of the multi-task fusion operation mode is the pipeline operation composed of corresponding calculation units.

3. The RDMA offloading method for database query tasks based on multi-stage pipeline according to claim 1, characterized in that: The single-task operation mode is a single sorting operation P, aggregation operation J, filtering operation G or connection operation L, and the multi-task fusion operation mode includes filtering operation G+sorting operation P, filtering operation G+connection operation L, filtering operation G+sorting operation P+connection operation L, filtering operation G+aggregation operation J, sorting operation P+aggregation operation J and sorting operation P+connection operation L.

4. The RDMA offloading method for database query tasks based on multi-stage pipeline according to claim 1, characterized in that: The database query task is generated by the host node Host, and the database query task is issued as follows: by calling the kernel API function, the parameters of the relevant task type and the destination computing unit are transmitted to the computing unit through the PCIE bus, and at the same time, the CPU calls the XRT API to cache the data to be processed to the memory unit M via the PCIE DMA. At the same time, the corresponding computing units are deployed in different computing units according to the task type, and a start signal is generated.

5. The RDMA offloading method for database query tasks based on multi-stage pipeline as claimed in claim 3, characterized in that: When the single-task operation mode is the connection operation L, the computing node deploys the connection computing unit, and the memory unit M4 is divided into multiple storage areas, namely M4-1, M4-2, M4-3, and M4-4, where M4-1 and M4-2 store the inner table data and the outer table data respectively.

6. The RDMA offloading method for database query tasks based on multi-stage pipeline according to claim 1, characterized in that: When the database issues a basic operation task in the single-task operation mode in a query, the computing system schedules the corresponding computing unit to complete the operation and returns the result to the corresponding memory unit and finally to the CPU memory of the host node.

7. A database query task RDMA offloading system based on a multi-stage pipeline, characterized in that: include: Computing node architecture building module, used to build a distributed computing node architecture centered on RDMA network units; The database query task offloading module is used to directly complete the forwarding offloading of database query task data through the RDMA network and the memory units of different computing units without CPU intervention based on the distributed computing node architecture centered on the RDMA network unit; The calculation of database query tasks includes single-task operation mode and multi-task fusion operation mode. When the task issued is in multi-task fusion operation mode, the corresponding computing unit is selected according to the basic operations contained in the multi-task to form a pipeline operation. The intermediate calculation results are directly forwarded to the memory unit corresponding to the computing unit, and then directly cached in the memory unit of the next-level computing unit via the RDMA network. The next-level computing unit reads the intermediate results in the memory unit to continue the calculation, and so on until the final result is returned to the CPU memory of the host node.

8. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the multi-stage pipeline-based RDMA offloading method for database query tasks described in any one of claims 1 to 6 is implemented.

9. A non-transitory computer-readable storage medium, characterized in that: The non-transitory computer-readable storage medium is used to store computer instructions. When the computer instructions are executed by the processor, the multi-stage pipeline-based database query task RDMA offloading method as described in any one of claims 1-6 is implemented.

10. An electronic device, characterized in that: include: A processor, a memory and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to perform the RDMA offloading method for database query tasks based on a multi-stage pipeline as described in any one of claims 1-6.

Images