الفهرس 
Background and Related Work
GPU Computing HistoryOnly 14 pages are availabe for public view
 |  | from | 127 |  |  |
| | | from | 127 | | |
Abstract
Data is evolving and the number of existing data sources is vastly growing. This necessitates the need for efficient and effective methods to store and process such huge amount of data. One of such methods that achieves
significant speed-ups at a small cost is employing a co-processor such as
GPU. In a previous work, the researchers demonstrated that GPU-acceleration cannot achieve significant speedups if the data has to be fetched from the disk, because of the IO bottleneck. The current research focuses on how to reduce the execution time of query processing and how to reduce the I/O transferring data cost during query processing. The GPU only improves performance of the data that is already available in the main memory. A recent research has shown that the GPU can be designed as an accelerator for individual database operations such as sort and joins, while another approach implements a database optimizer taking into account query plans with different operator placements, which increases the
size of the optimization space. However, existing approaches and database systems making use of GPU
acceleration have a common problem; they need to decide for each database
operator, on which heterogeneous processor it should be executed.Typically, each system has a set of analytical cost models, which model the performance behavior of an operator on a particular processor (e.g., CPUs or GPUs). To deal with this problem, in this thesis, we introduce an adaptive query processing approach which executes with SQL optimization plan using a Hybrid Query Processing. In particular, the approach is to automatically manage the distribution of the query parts to the CPU and GPU. The proposed approach parses SQL statements to generate subquery(part) statements. The approach then automatically manages the distribution
of the query parts to be executed on the CPU or in parallel on the GPU. To
carry out, we make use of the two dimensional structure of the CUDA grid
and thread model to perform query execution. We developed and implemented the proposed approach on a SQL server
database developed using the .Net framework instead of working under
Linux to exploit working with GPU on different environment that is used by
many developers around the world. To validate the performance of the approach, we conducted an extensive set of experiments utilizing workloads
with different characteristics. The results demonstrate that a GPU coprocessor can significantly improve performance on large data processing
tasks. In our work, the final result for implementation of the SQL queries,
that are given at the end of thesis in Appendix A, is that GPU execution is
4x-39x faster than CPU execution because in our work, we implement the
nested loop of the CROSS JOIN operation using both CPU and GPU instead of using only GPU or only CPU. But the total final result for implementation the SQL statements using virtual machine model proved that GPU execution is 2x - 5x faster than multi-core CPU execution when all in data and results transfer to and from the GPU are included.