high performance computing on graphics processing units: hgpu.org

In this dissertation we had the aim of utilizing GPU technology in order to optimize and improve on the problem of EST clustering. Extensive research on this cross-disciplinary approach was required before even considering such an approach. It was found that though this line of research has not received significant attention, there are significant gains that can be made through a project that utilizes GPU computing for bioinformatics problems. GPU programming differs from classical CPU programming in significant ways so a familiarity with the CUDA API is needed in order to achieve the performance goals. Understanding of the various types of parallelism and memory provided by GPUs is essential to optimizing the execution of a CUDA application. The metrics for performance and sensitivity measurement is important to consider for fair comparison between different platforms. The details and goals for the project is defined and expectations used for a measurement of success is defined before the application is implemented. EST clustering is a wide field with no single correct algorithm or implementation. For this reason extensive research had to be done in order to identify potential algorithms that this project will utilize. Each of the proposed algorithms are analysed for suitability for the GPU platform with their weaknesses and strengths identified. Most had to be discarded due to the limited scope of the project, but suitable algorithms for porting were found. Implementation involved a lot of learning, adapting, and a few surprises, but eventually a program was completed that met the goals of the project.