Publication Date


Date of Final Oral Examination (Defense)


Type of Culminating Activity


Degree Title

Master of Science in Computer Science


Computer Science

Major Advisor

Elena A. Sherman, Ph.D.

Major Advisor

İnanç Şenocak, Ph.D.


Steven M. Cutchin, Ph.D.


The research presented in this thesis investigates parallel implementations of the Fast Sweeping Method (FSM) for Graphics Processing Unit (GPU)-based computational plat forms and proposes a new parallel algorithm for distributed computing platforms with accelerators. Hardware accelerators such as GPUs and co-processors have emerged as general- purpose processors in today’s high performance computing (HPC) platforms, thereby increasing platforms’ performance capabilities. This trend has allowed greater parallelism and substantial acceleration of scientific simulation software. In order to leverage the power of new HPC platforms, scientific applications must be written in specific lower-level programming languages, which used to be platform specific. Newer programming models such as OpenACC simplifies implementation and assures portability of applications to run across GPUs from different vendors and multi-core processors.

The distance field is a representation of a surface geometry or shape required by many algorithms within the areas of computer graphics, visualization, computational fluid dynamics and more. It can be calculated by solving the eikonal equation using the FSM. The parallel FSMs explored in this thesis have not been implemented on GPU platforms and do not scale to a large problem size. This thesis addresses this problem by designing a parallel algorithm that utilizes a domain decomposition strategy for multi-accelerated distributed platforms. The proposed algorithm applies first coarse grain parallelism using MPI to distribute subdomains across multiple nodes and then fine grain parallelism to optimize performance by utilizing accelerators. The results of the parallel implementations of FSM for GPU-based platforms showed speedup greater than 20× compared to the serial version for some problems and the newly developed parallel algorithm eliminates the limitation of current algorithms to solve large memory problems with comparable runtime efficiency.