Research On Hardware Acceleration Of Database System

Driven by big data and cloud computing,the query performance of database system is also developing in the direction of high throughput and low latency.In the fields of biomedical engineering,signal processing,artificial intelligence,etc.,a large amounts of data such as biomedical data,real-time signal and video data exist,the queries of these big data put forward higher requirements for the performance of database systems.Therefore,it is of great significance to improve the performance of database system for its wide application in the context of big data.In this thesis,from the perspective of hardware acceleration and taken relational database system as the research object,I explore the techniques on acceleration of database systems with an FPGA-based accelerator.Firstly,considering the limited processing power of the acceleration framework with a single acceleration node,a database distributed acceleration framework which is based on Infiniband network is proposed.The acceleration framework,composed of the host nodes which generate accelerated tasks,the acceleration nodes which execute specific accelerated tasks and the infiniband facilities,which are used as the interconnect between the two kinds of nodes,can realize collaborative parallel acceleration of accelerated tasks through multiple acceleration nodes.Based on the proposed acceleration framework,a database-oriented distributed acceleration system is designed which comprises three kinds of components: the acceleration protocol,the software acceleration subsystem and the hardware acceleration subsystem.The acceleration protocol defines the whole process by which the host node offloads the acceleration requests to the hardware accelerators through the acceleration library interface layer,the acceleration library abstract layer,the acceleration library management layer and the acceleration library driver layer;The software acceleration mainly completes the registration of acceleration functions,the initialization of the acceleration parameters,the memory allocation of the buffer in the acceleration node,and the data transmission between the host node and the acceleration node according to the acceleration protocol;The hardware acceleration subsystem is responsible for the execution of the accelerated task.It completes different accelerated tasks through the acceleration execution unit-acceleration IP.Secondly,in view of the lack of query acceleration of the ordinary database storage engines,this thesis proposed an storage engine that supports dynamic offloading and parallel hardware acceleration of query operations like filtering and projection.The storage engine offloads the filtering and projection operations to the hardware accelerator located between the storage device and the host node,thus realizing the fusion of data transmission and query acceleration.In addition,through the hierarchical design of the storage engine which includes the table manager,the acceleration node manager,and the acceleration node driver,the content of the table is stored in the storage device managed by each acceleration node in a distributed manner.Thus,a single query operation can be divided and allocated to each acceleration node and executed in parallel.Thirdly,considering the shortcomings of sorting and join operations in conventional database systems in terms of throughput and latency,this thesis completed two algorithm structures with high performance and capable of handling sorting and join queries respectively.· Aiming at the sorting operation,this thesis proposed a high-performance algorithm architecture for top-k sorting.The architecture has the characteristics of supporting variable-length sequences and continuous sequences.Through the setting of the pa-rameter data parallelism P,the throughput and delay of the sorting architecture can be adjusted.· For join operation,this thesis proposed a join algorithm based on a top-k sorter.The algorithm applies a filter to remove tuples having no potential matches with the aim of reducing the number of comparisons during join operations between two tables.Based on this algorithm,a modular pipeline join architecture is designed.Not only high throughput is achieved,but also the resource consumption of the architecture is unaffected by the table size such that a join operation between tables with arbitrary sizes is supported.Finally,an experimental platform is built to verify the relevant techniques.The experiments demonstrated that the database acceleration system which is based on the distributed acceleration framework can feasibly speed up the database operations of sorting and join.In addition,the proposed storage engine can support queries offloading and speed up database operations of filtering and projection by means of a FPGA-based hardware accelerator.