Relational database join connection engine system based on FPGA

By designing a relational database join connection engine system on FPGA, the hash table construction and detection module is used to resolve hash conflicts, and the join connection of variable width data two-dimensional tables is supported, which solves the current technology of join connection speed on FPGA. The problem of limitations enables efficient join connection processing and flexible application scenarios.

CN119088864APending Publication Date: 2024-12-06山东浪潮数据库技术有限公司
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411068248.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-06
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

The processing speed of the existing relational database join connection operation on the FPGA is affected by the hash collision detection search process, and does not support join connections of variable-width data two-dimensional tables, which limits the application scope.

Method used

设计一种基于FPGA的关系型数据库join连接引擎系统,包括数据库总控服务器和数据join连接加速卡,利用FPGA芯片进行哈希表构建和探测,通过PCIe接口实现高速数据交换,支持多种哈希函数和并行处理,解决哈希冲突问题,并支持可变宽度数据二维表的join连接。

Benefits of technology

It significantly improves the efficiency of join connection processing, supports complex high-speed join connection query, reduces time loss and is flexible in application. The experimental results show that the time loss is reduced by 86.8% compared with the traditional solution under the same conditions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119088864A_ABST
    Figure CN119088864A_ABST
Patent Text Reader

Abstract

The invention provides a relational database join connection engine system based on an FPGA, and belongs to the field of database application design, the relational database join connection engine system comprises a database master control server, and a server CPU is used for processing a join connection query request task of a user; the data join connection acceleration card is responsible for executing a high-speed join connection calculation task of the target data two-dimensional table; and the PCIe interface is used for connecting the database general control server and the data join connection acceleration card. The join connection processing efficiency is improved to the greatest extent, and meanwhile, the time loss is also reduced.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of database application design, and in particular to a relational database join connection engine system based on FPGA. Background Art

[0002] A relational database is a data record storage system that uses a relational model to organize data. It stores data in the form of rows and columns and supports all-round storage, search, and analysis of data. It is widely used in business, industry, smart home, and healthcare. Users can discover information hidden in data sets by performing scientific and reasonable aggregate queries on relational databases. Performing join queries on relational databases will trigger various two-dimensional table merge operations, realize the splicing and union of two-dimensional tables, and facilitate data integration.

[0003] Most of the existing research on relational database aggregation query engines only implements ordinary join operations on FPGA, that is, based on the multi-dimensional join architecture of R-Tree, using "hierarchical index" to parallelize tasks and accurately perform join matching. However, due to the large amount of data involved in joins, hash conflicts are inevitable, and the speed advantage of processing data on FPGA is seriously affected by the detection and search process caused by a large number of hash conflicts. In addition, existing research does not support joins of variable-width data two-dimensional tables, which greatly affects the application scope of join operations. For this reason, a new FPGA-based relational database join engine urgently needs to be explored in depth by researchers in related fields. Summary of the invention

[0004] In order to solve the above problems, the present invention analyzes the key points of the relational database join connection processing engine to execute different functions and data processing specifications, and proposes an FPGA-based relational database join connection engine system to meet the complex high-speed join connection query requirements of the relational database.

[0005] The technical solution of the present invention is:

[0006] A relational database join connection engine system based on FPGA, comprising:

[0007] The database master control server, whose CPU is used to process the user's join query request task;

[0008] Data join connection accelerator card, responsible for executing high-speed join connection calculation tasks of the target data two-dimensional table;

[0009] PCIe interface, used to connect to the database master control server and data join connection accelerator card.

[0010] Furthermore,

[0011] The database master control server is connected to the database user terminal to complete the user demand analysis and data two-dimensional table reading and processing work;

[0012] The data join connection acceleration card includes an FPGA chip, a PCIe interface, a crystal oscillator module, a power module, a debugging interface, and necessary resistors and capacitors. The acceleration card is responsible for processing the aggregation processing tasks received from the database master server and outputting the processing results to the server user end in a timely manner.

[0013] Going further,

[0014] The specific process of processing the aggregation task is as follows:

[0015] a. After receiving the user's join connection processing instruction, the database host CPU performs instruction parsing and optimization, and determines the corresponding storage partition and address of the two-dimensional data table involved in the join connection operation in the storage array;

[0016] b. The CPU reads the two-dimensional table of data involved in the proposed join operation from the corresponding memory address;

[0017] c. Send the join instruction and the extracted two-dimensional table of data involved in the join operation to the FPGA chip of the join accelerator card through PCIe;

[0018] d. The FPGA chip adjusts the join processing frequency according to the required join instructions and performs join calculations on the read two-dimensional data table;

[0019] d. The FPGA chip packages the join calculation results and sends them to the database master control server through the PCIe interface. After completion, it sends a transmission completion signal to the CPU;

[0020] e. After receiving the completion signal, the CPU of the database master control server sends the result to the user terminal to complete the single join calculation.

[0021] in,

[0022] The FPGA chip includes a DDR4 read-write control module, an AXI polling module, a hash table building module, a CMM management module, a hash table detection module, a BRAM storage, a hash table reset module, a connection result sending module, an HBM memory read-write module, and a logic management module to collaboratively complete the join connection processing task.

[0023] The join query acceleration kernel workflow is as follows:

[0024] (1) The logic management module obtains configuration information from the host computer and stores the row length, number of rows, column width, and key value column position information of the internal and external tables, providing a parameter basis for the subsequent construction and detection stages;

[0025] (2) During the construction or detection phase, the AXI polling module reads the inner and outer table data to be joined from DDR4;

[0026] (3) The hash table construction module uses the hash function to discretely store the data into the global memory based on the internal table data to construct a hash table;

[0027] (4) The CMM management module applies for a separate storage space to store the row data that cannot be inserted into the hash table to solve the hash conflict problem;

[0028] (5) The hash table detection module reads the external data during the detection phase of the hash connection and uses the hash address to detect the hash table to find matching data;

[0029] (6) After the detection is completed, the connection result sending module re-converts the data format according to the matching result, and returns the re-joined two-dimensional table data to the DDR4 global memory;

[0030] (7) When a hash connection is completed, the hash table reset module resets all hash tables, releases hash table resources, and waits for the next hash connection task.

[0031] Going further,

[0032] For the application scenario of the connection result, the hash connection hardware based on the MV-PHJ algorithm is used for implementation. The specific steps are as follows:

[0033] The XDMA controller reads the internal table row data from the accelerator card DDR4, and the parsing module extracts the column connection attribute. If it is a string, the connection attribute is converted into an integer key value through string hash operation. If it is a Boolean or integer, no processing is required;

[0034] Then, the connection attribute type type, key value key, and row length rlength are written into the Info Buffer of the corresponding Virtual Channel in the data structure of {type, rlength, key}, and the internal table elements are written into the Data Buffer;

[0035] MV-PHJ constructs four VC channels, each of which is designed with parallel pipelines. If the inner table element cannot be inserted into the hash table by traversing the four channels, the element request is inserted into the CMM memory, that is, the overhead data is stored in the I-BRAM and the inner table data is stored in the D-BRAM, completing a single hash construction process.

[0036] Going further,

[0037] After the MV-PHJ construction phase is completed, the MV-PHJ detection method is executed. The XDMA controller reads the row data of the outer table from DDR4, and the parsing module extracts the column connection attribute. If it is a string, the connection attribute is converted into an integer key value key by string hash operation. If it is a Boolean or integer, no processing is required. Then the connection attribute type type, key value ke), and row length rlength are written into the InfoBuffer of the corresponding Virtual Channel channel in the data structure of {type, rlength, key}. In each VC channel, the hash table is searched according to the index address of the key hash calculation result, and compared with the key value in the returned overhead data; when the key value is the same, it means that the connection is successful, and the outer table data and the corresponding inner table data in the hash table Bank0 are read to form the connection result output. If the key value is different, the overhead data and the outer table data are inserted into the next VC channel to continue detection; when no matching object is found after traversing the VC channel and CMM memory, the outer table data is discarded and this detection ends.

[0038] Going further,

[0039] Select the following hash functions to perform hash operations on the data to be connected:

[0040] 1) CRC check method, using CRC32-D32 and CRC32-D64 algorithms to construct hash functions, the expression is as follows:

[0041] H0(x)=crc(A,Key)&Mask0 (1)

[0042] Where A is the generating polynomial, Key is the key value, and Mask is 0x3fffff. Add the result of the crc function to the mask, and take the lower 22 bits as the hash address;

[0043] 2) Square middle method, use the square middle algorithm to construct a hash function, the expression is as follows:

[0044] H1(x)=square(Key,Key)&Mask1 (2)

[0045] Among them, Key is the input key value, Mask is 0x7ffffe0, and after squaring, the middle 22 bits of the result are taken as the hash address;

[0046] 3) Address accumulation method, using direct addressing algorithm to construct hash function, the expression is as follows:

[0047] H2(x)=(Key[31:22]==0)? Key[21:0]:(Key[21:0]+Key[31:22]) (3)

[0048] Among them, Key is the input key value. When the upper 10 bits are false, the lower 22 bits are taken as the hash address; when the upper 10 bits are true, the sum of them is taken as the hash address.

[0049] When the connection attribute is a string, the string needs to be converted into a 64-bit integer through a hardware-implemented string hash operation, and then the hash function is executed to obtain the hash table index address. The algorithm expression is as follows:

[0050] H(0)=String(0)(4)

[0051] H(i)=H(i-1)·P+String(i) (5)

[0052] Among them, String represents a string element, P represents a constant, and i represents a string label.

[0053] The beneficial effects of the present invention are

[0054] The rapid growth of storage capacity has made join connection query a major challenge facing modern relational databases. The present invention proposes a relational database join connection engine based on FPGA. The relational database join connection engine based on FPGA not only supports high-speed join connection processing operations for different types of data, but also supports the execution of join connection processing for variable-width two-dimensional data tables, supports FPGA core reuse and parallel processing, and maximizes the efficiency of join connection processing. In addition, the relational database join connection engine based on FPGA can also be reconfigured according to the actual needs of join connection queries, and can support a variety of different join connection operation updates, and is flexible in application. The database query experiment on the client of Inspur Yingxin Server (NF5468M6) shows that when querying the same content, the FPGA-based database join connection engine solution proposed by the present invention reduces the time loss by up to 86.8% compared with the traditional solution of aggregate query processing through CPU+FPGA. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a schematic diagram of the hardware architecture of the present invention;

[0056] Figure 2 It is a schematic diagram of the FPGA module structure of the relational database join connection engine;

[0057] Figure 3 It is a schematic diagram of the construction algorithm of MV-PHJ;

[0058] Figure 4 It is a schematic diagram of the MV-PHJ detection algorithm. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0060] The present invention analyzes the key points of relational database join connection processing engine to execute different functions and data processing specifications, and proposes a relational database join connection engine system based on FPGA to meet the complex high-speed join connection query requirements of relational database, including the following aspects:

[0061] 1. Hardware architecture and working mode of FPGA-based relational database join engine

[0062] Figure 1 The hardware architecture and working mode of the present invention are demonstrated. The system includes a database master control server (the server CPU is used to process the user's join connection query request task), a data join connection acceleration card (the acceleration card is responsible for executing the high-speed join connection calculation task of the target data two-dimensional table) and a PCIe interface connecting the two.

[0063] Among them, the database master control server is the database function core. This device connects to the database user terminal and can complete the user demand analysis and data two-dimensional table reading and processing work; the data join connection acceleration card includes an FPGA chip (the FPGA chip deploys all the cores required to execute the data join connection), a PCIe interface, a crystal oscillator module, a power module, a debugging interface, and necessary resistors and capacitors. The acceleration card is responsible for processing the aggregation processing tasks received from the database master control server and outputting the processing results to the server user end in a timely manner; the data join connection acceleration card is connected to the CPU of the master control server through the PCIe high-speed transmission interface to achieve high-speed data exchange;

[0064] The single data aggregation calculation process of the column-oriented data aggregation query acceleration design can be briefly described as follows:

[0065] a. After receiving the user's join connection processing instruction, the database host CPU performs instruction parsing and optimization, and determines the corresponding storage partition and address of the two-dimensional data table involved in the join connection operation in the storage array;

[0066] b. The CPU reads the two-dimensional table of data involved in the proposed join operation from the corresponding memory address;

[0067] c. Send the join instruction and the extracted two-dimensional table of data involved in the join operation to the FPGA chip of the join accelerator card through PCIe;

[0068] d. The FPGA chip adjusts the appropriate join processing frequency (maximum efficiency principle) according to the required join instructions, and performs join calculations on the read two-dimensional data table;

[0069] d. The FPGA chip packages the join calculation results and sends them to the database master control server through the PCIe interface. After completion, it sends a transmission completion signal to the CPU;

[0070] e. After receiving the completion signal, the CPU of the database master control server sends the result to the user terminal to complete the single join calculation.

[0071] 2. FPGA module design of relational database join connection engine

[0072] The specific FPGA module design of the relational database join connection engine is composed of Figure 2 The architecture includes DDR4 read / write control module (DDR4 Control), AXI polling module (AXI pooling), hash table construction module (Hash table construction), CMM (Conflict Management Memory) management module (CMM management), hash table detection module (Hash table detection), BRAM storage (BRAM), hash table reset module (Hash table reset), connection result sending module (Join result TX), HBM memory read / write module (HBM RW unit), and necessary logical management (Logical management) modules to complete the join connection processing task. In order to maximize the performance of hash connection and achieve a balance between the computational overhead and acceleration effect of FPGA hash algorithm calculation, this work adopts a serial-parallel pipeline method to construct and detect the hash table.

[0073] The specific join query acceleration kernel logic is implemented as follows:

[0074] (1) The logic management module obtains configuration information from the host computer and stores information such as row length, number of rows, column width, key value column position, etc. of the internal and external tables, providing a parameter basis for the subsequent construction and detection stages.

[0075] (2) During the construction or detection phase, the AXI polling module reads the inner and outer table data to be joined from DDR4.

[0076] (3) The hash table construction module uses the hash function to discretely store the data into the global memory based on the internal table data to build a hash table.

[0077] (4) The CMM management module applies for a separate storage space to store the row data that cannot be inserted into the hash table, effectively solving the hash conflict problem.

[0078] (5) The hash table detection module reads the external data during the detection phase of the hash connection and uses the hash address to detect the hash table to find matching data.

[0079] (6) After completing the detection, the connection result module reconverts the data format according to the matching results and returns the rejoined two-dimensional table data to the DDR4 global memory.

[0080] (7) When a hash connection is completed, the hash table reset module resets all hash tables, releases hash table resources, and waits for the next hash connection task.

[0081] 3. Design of MV-PHJ hash join algorithm

[0082] Aiming at the application scenario where the connection result is unique, a hash connection hardware implementation based on the MV-PHJ algorithm is proposed. This strategy makes up for the shortcomings of unlinked list parallel hash connection and supports FPGA-based string connection properties; the hash table is constructed through the global memory HBM (High Bandwidth Memory), and the tuple data and overhead data are cached independently, which can flexibly adapt to different data types.

[0083] The construction algorithm of MV-PHJ is as follows Figure 3As shown in the figure, the XDMA controller reads the internal table row data from the accelerator card DDR4, and the column connection attribute is extracted by the parsing module. If it is a string, the connection attribute is converted into an integer key value by string hash operation. If it is a Boolean or integer, no processing is required. Then the connection attribute type (type), key value (key), and row length (rlength) are written into the Info Buffer of the corresponding Virtual Channel (VC) channel in the data structure of {type, rlength, key}, and the internal table elements are written into the Data Buffer. MV-PHJ constructs four VC channels, and each channel is designed in parallel pipeline. Taking VC0 as an example, when the overhead data of the Info Buffer using the global memory HBM of the FPGA is detected, the Key value is extracted as the input of the hash function h1 of VC0 to obtain the index address of hash table 1. Then, the overhead hash table Bank1 is read according to the index address and it is determined whether the key value, type value and Valid are valid, and whether a hash conflict occurs, where Valid indicates whether the current index address is occupied. If no hash conflict occurs, the overhead hash table Bank1 pointed to by the current index address is written in the data format of {offset, type, rlength, key, valid}, and the internal table data is written to the data hash table Bank0 in sequence from address 0, and then the offset is recalculated as the starting address for the next write to the data hash table Bank0, where the initial value of offset is 0 and needs to be recalculated for each construction. If a conflict occurs, the overhead data and internal table data are written to the reinsertion buffer of VC1, and the same construction process continues. If the internal table element cannot be inserted into the hash table after traversing the four channels, the element request is inserted into the CMM memory (the overhead data is stored in I-BRAM and the internal table data is stored in D-BRAM) to complete the single hash construction process.

[0084] After the MV-PHJ construction phase is completed, the following steps are executed: Figure 4The MV-PHJ detection algorithm shown in the figure. As in the construction phase, the XDMA controller reads the row data of the outer table from DDR4, and the parsing module extracts the column connection attribute. If it is a string, the connection attribute is converted into an integer key value (key) by string hash operation. If it is a Boolean or integer, no processing is required. Then the connection attribute type (type), key value (key), and row length (rlength) are written into the Info Buffer of the corresponding VC (Virtual Channel) channel in the data structure of {type, rlength, key}. In each VC channel, the hash calculation works similarly to the construction phase. The hash table is searched according to the index address of the key hash calculation result and compared with the key value in the returned overhead data. When the key value is the same, it means that the connection is successful, and the outer table data and the corresponding inner table data in the hash table Bank0 are read to form the connection result output. If the key value is different, the overhead data and the outer table data are inserted into the next VC channel to continue detection. When no matching object is found after traversing the VC channel and CMM memory, the outer table data is discarded and this detection ends.

[0085] 4. Hash function design

[0086] This work uses the global memory resource HBM of FPGA to build a hash table, and uses the BRAM memory on the FPGA chip to build a CMM. Each VC is adapted to a hash function, and the parallel virtual channel can be flexibly expanded according to the hardware resources. It is very important to reasonably design the hash function of the VC channel so that the elements of the inner table are evenly distributed to the hash table. The present invention selects the following hash functions to perform the hash operation of the data to be connected.

[0087] CRC check method uses CRC32-D32 and CRC32-D64 algorithms to construct hash functions. The expression is as follows:

[0088] H0(x)=crc(A,Key)&Mask0 (1)

[0089] Where A is the generator polynomial, Key is the key value, and Mask is 0x3fffff. Add the result of the crc function to the mask, and take the lower 22 bits as the hash address.

[0090] The square middle method uses the square middle algorithm to construct a hash function. The expression is as follows:

[0091] H1(x)=square(Key,Key)&Mask1 (2)

[0092] Among them, Key is the input key value, Mask is 0x7ffffe0, and after squaring, the middle 22 bits of the result are taken as the hash address.

[0093] The address accumulation method uses a direct addressing algorithm to construct a hash function. The expression is as follows:

[0094] H2(x)=(Key[31:22]==0)? Key[21:0]:(Key[21:0]+Key[31:22]) (3)

[0095] Among them, Key is the input key value. When the upper 10 bits are false, the lower 22 bits are taken as the hash address; when the upper 10 bits are true, the sum of them is taken as the hash address.

[0096] In addition, when the connection attribute is a string, the string needs to be converted into a 64-bit integer through hardware-implemented string hashing, and then the hash function is executed to obtain the hash table index address. The algorithm expression is as follows:

[0097] H(0)=String(0) (4)

[0098] H(i)=H(i-1)·P+String(i) (5)

[0099] Among them, String represents a string element, P represents a constant, and i represents a string label.

[0100] 5. Test methods and results

[0101] In order to verify the reliability of the FPGA-based relational database join connection engine of the present invention. Inspur Yingxin server (NF5468M6) is selected as the host node, and Inspur F37X data acceleration card is used as the carrier to implement the join connection query engine, and KaiwuDB relational database is selected as the database. The FPGA-based database join connection query scheme proposed by the present invention and the traditional CPU+FPGA connection query scheme are used to perform join connection query tests on data packets of different sizes under the same relational data and query conditions, and the time required for each group of experimental queries is recorded.

[0102] The specific experimental query performance results are shown in the following table:

[0103]

[0104]

[0105] The result comparison shows that when executing join connections with the same content, the time required for querying the FPGA-based relational database join connection engine proposed by the present invention is reduced by 86.8% compared with the processing time of the traditional solution, which greatly speeds up the efficiency of relational database join connection queries and meets the real-time requirements of query processing operations.

[0106] The above description is only a preferred embodiment of the present invention, which is only used to illustrate the technical solution of the present invention, and is not used to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims

1. A relational database join connection engine system based on FPGA, characterized in that: include The database master control server, whose CPU is used to process the user's join query request task; Data join connection accelerator card, responsible for executing high-speed join connection calculation tasks of the target data two-dimensional table; PCIe interface, used to connect to the database master control server and data join connection accelerator card.

2. The system according to claim 1, characterized in that The database master control server connects to the database user terminal to complete the user demand analysis and data two-dimensional table reading and processing; The data join connection acceleration card includes an FPGA chip, a PCIe interface, a crystal oscillator module, a power module, a debugging interface, and resistors and capacitors. The acceleration card is responsible for processing the aggregation processing tasks received from the database master server and outputting the processing results to the server user end in a timely manner.

3. The system according to claim 2, characterized in that The specific process of processing the aggregation task is as follows: a. After receiving the user's join connection processing instruction, the database host CPU performs instruction parsing and optimization, and determines the corresponding storage partition and address of the two-dimensional data table involved in the join connection operation in the storage array; b. The CPU reads the two-dimensional table of data involved in the proposed join operation from the corresponding memory address; c. Send the join instruction and the extracted two-dimensional table of data involved in the join operation to the FPGA chip of the join accelerator card through PCIe; d. The FPGA chip adjusts the join processing frequency according to the required join instructions and performs join calculations on the read two-dimensional data table; d. The FPGA chip packages the join calculation results and sends them to the database master control server through the PCIe interface. After completion, it sends a transmission completion signal to the CPU; e. After receiving the completion signal, the CPU of the database master control server sends the result to the user terminal to complete the single join calculation.

4. The system according to claim 3, characterized in that The FPGA chip includes a DDR4 read-write control module, an AXI polling module, a hash table building module, a CMM management module, a hash table detection module, a BRAM storage, a hash table reset module, a connection result sending module, an HBM memory read-write module, and a logic management module to collaboratively complete the join connection processing task.

5. The system according to claim 4, characterized in that The join query acceleration kernel workflow is as follows: (1) The logic management module obtains configuration information from the host computer and stores the row length, number of rows, column width, and key value column position information of the internal and external tables, providing a parameter basis for the subsequent construction and detection stages; (2) During the construction or detection phase, the AXI polling module reads the inner and outer table data to be joined from DDR4; (3) The hash table construction module uses the hash function to discretely store the data into the global memory based on the internal table data to construct a hash table; (4) The CMM management module applies for a separate storage space to store the row data that cannot be inserted into the hash table to solve the hash conflict problem; (5) The hash table detection module reads the external data during the detection phase of the hash connection and uses the hash address to detect the hash table to find matching data; (6) After the detection is completed, the connection result sending module re-converts the data format according to the matching result, and returns the re-joined two-dimensional table data to the DDR4 global memory; (7) When a hash connection is completed, the hash table reset module resets all hash tables, releases hash table resources, and waits for the next hash connection task.

6. The system according to claim 5, characterized in that For the application scenario of the connection result, the hash connection hardware based on the MV-PHJ algorithm is used for implementation. The specific steps are as follows: The XDMA controller reads the internal table row data from the accelerator card DDR4, and the parsing module extracts the column connection attribute. If it is a string, the connection attribute is converted into an integer key value through string hash operation. If it is a Boolean or integer, no processing is required; Then, the connection attribute type type, key value key, and row length rlength are written into the Info Buffer of the corresponding Virtual Channel in the data structure of {type, rlength, key}, and the internal table elements are written into the Data Buffer; MV-PHJ constructs four VC channels, each of which is designed with parallel pipelines. If the inner table element cannot be inserted into the hash table by traversing the four channels, the element request is inserted into the CMM memory, that is, the overhead data is stored in the I-BRAM and the inner table data is stored in the D-BRAM, completing a single hash construction process.

7. The system according to claim 6, characterized in that After the MV-PHJ construction phase is completed, the MV-PHJ detection method is started. The XDMA controller reads the row data of the outer table from DDR4, and the parsing module extracts the column connection attributes. If it is a string, the connection attribute is converted into an integer key value key through a string hash operation. If it is a Boolean or integer, no processing is required. Then the connection attribute type type, key value ke), and row length rlength are written to the InfoBuffer of the corresponding Virtual Channel channel in the data structure of {type, rlength, key}. In each VC channel, the hash table is searched according to the index address of the key hash calculation result, and compared with the key value in the returned overhead data; when the key value is the same, it means that the connection is successful, and the outer table data and the corresponding inner table data in the hash table Bank0 are read to form the connection result output. If the key value is different, the overhead data and the outer table data are inserted into the next VC channel to continue detection; If no matching object is found after traversing the VC channel and the CMM memory, the appearance data is discarded and the detection ends.

8. The system according to claim 7, characterized in that Select the following hash functions to perform hash operations on the data to be connected: 1) CRC check method, using CRC32-D32 and CRC32-D64 algorithms to construct hash functions, the expression is as follows: H0(x)=crc(A,Key)&Mask0 (1) Where A is the generating polynomial, Key is the key value, and Mask is 0x3fffff; add the result of the crc function to the mask, and take the lower 22 bits as the hash address; 2) Square middle method, use the square middle algorithm to construct a hash function, the expression is as follows: H1(x)=square(Key,Key)&Mask1 (2) Among them, Key is the input key value, Mask is 0x7ffffe0, and after squaring, the middle 22 bits of the result are taken as the hash address; 3) Address accumulation method, using direct addressing algorithm to construct hash function, the expression is as follows: H2(x)=(Key[31:22]==0)? Key[21:0]:(Key[21:0]+Key[31:22]) (3) Among them, Key is the input key value. When the upper 10 bits are false, the lower 22 bits are taken as the hash address; when the upper 10 bits are true, the sum of them is taken as the hash address.

9. The system according to claim 8, characterized in that When the connection attribute is a string, the string needs to be converted into a 64-bit integer through a hardware-implemented string hash operation, and then the hash function is executed to obtain the hash table index address. The algorithm expression is as follows: H(0)=String(0)(4) H(i)=H(i-1)·P+String(i)(5) Among them, String represents a string element, P represents a constant, and i represents a string label.