Optimization Methods For Complex SQL Query Based On GPU

With the development of information technology, the scale of the data indatabase is becoming more and more larger, the characteristics of the data are: largeramount, many data types, low density of value. In this background, the querieschange from traditional single dimension simple query to multi-dimensionalcomplex query. As an important means of analyzing the database data, complexquery plays an important role in processing period. By execute queries,decision-makers can get the information they concerned quickly. Also usingtraditional method to extract, store and analysis mass data is becoming increasinglydifficult to obtain real-time results. moreover, it also restrict the decision-making ofthe enterprise.In order to improve the analysis speed on large scale data, this paper proposes acolumn-oriented storage model for accerlerate database query speed and also astrategy for accelerate query by combing GPU parallel computing capability andcolumn-oriented database storage model. The main research contents of this paperare as follows:(1) First study basic theories about database complex query and GPU parallelcomputing model. Summed up the traditional database query optimizationtechniques. In-depth learning different storage strategy of different databases,especially focus on the study compression and storage strategies of differentdatabases;(2) Propose a sparse index based physical storage model, the model adoptsdivide strategy based on column-oriented storage model. Meanwhile we chooseFOR compression algorithm according to GPU features, and we achieve highparallel compression with GPU parallel computing capability. By using GPU, wecan greatly improve the compression speed compared to the compression speed onCPU;(3) Accelerate complex query operations by using GPU parallel computingtechnology and GPU query primitives, mainly focus on accelerate range query andgrouping query optimization, and the merge strategy on group results. Propose ascheduling strategy to solove the problem of long IO time during the period ofexperiments, By the scheduling strategy we can partly speed up the response speedof the query;(4) The experiment results shows the superiority of using GPU to acceleratecompression algorithm and query speed. We adopt TPC-H test data sets to analysisour query model and it proves that this model can obtain5-8times speedup compare to traditional database model under large data sets.