Design And Implementation Of Heterogeneous Computing Framework And Its Operators About Distributed Memory Database

With the advent of the era of big data,Online Analytical Processing(OLAP)has become the focus of research.The efficiency of mass data processing has become the key point of OLAP research.In order to improve execution efficiency,full memory operations are increasingly being used in OLAP databases.At the same time,CPU / GPU heterogeneous computing frameworks are used more and more widely.Applying heterogeneous computing to an OLAP database can greatly improve the execution efficiency of OLAP and reduce the cost of OLAP,which is of great significance in database science and engineering.At present,the main way to accelerate the database using GPU heterogeneous computing is to extract the computation-intensive operations in the database and then replace it with GPU parallel computing logic.This general approach cannot solve the problem of bus transmission bottlenecks.In this thesis,a complete set of heterogeneous acceleration schemes is designed to solve the bus transmission bottleneck of heterogeneous computing system of OLAP database,take GPU index as the main idea,and heterogeneous operator as the main ways.The goal of this thesis is to provide a heterogeneous acceleration scheme for OLAP distributed memory databases.The main work about this thesis is as follows:1.Designed the GPU indexing scheme as the basis of the full thesis.The GPU indexing scheme can not only make full use of the parallel computing capabilities of the GPU,but also greatly reduce the data transmission between memory and device memory,and solve the problem of bus transmission bottlenecks.2.A hybrid operator model is designed based on the GPU indexing scheme.In the mixed operator model,each operator has two parts: GPU approximate operator and CPU precise operator.The two parts cooperate with each other to complete the calculation task.This thesis also discusses in detail how to apply the mixed operator model to the key operators of the OLAP database.3.A scheduling model adapted to the GPU indexing scheme is designed.The scheduling model is divided into distributed global scheduling and single-node local scheduling.Hot data in a distributed cluster is migrated by the global scheduling,and indexes of local data is adjusted by the local scheduling.4.Solve some key problems of applying heterogeneous computing to the database,establish mathematical models for these problems and get many useful conclusions.These conclusions are directly applied to my system,and most of these are also of significance for other types of heterogeneous computing.Finally,the distributed memory database system with heterogeneous computing is tested.The test results show that the execution efficiency of this system is much higher than the efficiency of using the CPU alone.