Optimization For Parallel Data Operation On Gpu

In recent years, with the development and maturity of multicore technology, GPU(Graphics Processing Units) has increasingly become the research hot spot by its powerful parallel computing capacity. Researchers have been using GPU to accelerate various kinds of data operations, among which relational join is the most commonly used and most time-consuming operation.This paper has studied the three classic relational join algorithms (nested-loop join, sort-merge join and hash join), both serial and parallel version. Their parallel versions have been implemented on GPU and optimized according to their characteristics. At the meantime, due to the similarity between clusters and multicore system, this paper has migrated CMD(Coordinate Module Distribution) algorithm, originally designed for clusters, onto GPU and compared its performance with those two classic algorithms on GPU. We also compared their performances when processing tables with data skew. Analysis also has been done with OpenMP implementation. The contribution of this paper is the implementation and optimization of two classic relational join algorithms and migration of CMD algorithm. This paper compares the performance of those algorithms when facing data skew, which shows that hash-join outperforms the other two, with sort-merge join does the worst job. But when facing even-distributed data tables, CMD algorithm on GPU is inferior to those two classic algorithms. The GPU versions of sort-merge join and hash join has nearly the same performance with their OpenMP implementation...