This thesis investigates the challenges of parallel I/O in shared memory applications, particularly within the OpenMP framework. It explores the limitations of traditional I/O approaches in OpenMP and proposes a novel technique to enhance parallel I/O performance. The proposed method involves careful management of thread offsets and data chunk distribution to enable efficient parallel write operations. Extensive experiments were conducted on diverse high-performance computing systems, utilizing various storage devices such as NVMe SSDs, CephFS, SATA SSDs, and HDDs. The results demonstrate that the proposed parallel I/O technique significantly improves performance on high-throughput storage devices, achieving throughput close to the theoretical maximum. While the gains were less pronounced on slower HDDs due to their mechanical limitations, the technique still exhibited improvements in specific scenarios. This research contributes to the advancement of parallel I/O capabilities in shared memory environments, offering a promising solution to overcome I/O bottlenecks and enhance the overall efficiency of OpenMP applications. The findings highlight the potential for further development and integration of parallel I/O techniques into the OpenMP framework, paving the way for improved performance and scalability in future high-performance computing applications.

@mastersthesis{EOFPIISMAA24,
	author	 = {Shree Krishna Adhikari},
	title	 = {{Enhancing OpenMP for Parallelizing Input/Output in Shared Memory Applications}},
	advisors	 = {Michael Kuhn and Jannek Squar and Anna Fuchs},
	year	 = {2024},
	month	 = {09},
	school	 = {Universität Hamburg},
	howpublished	 = {{Online \url{https://wr.informatik.uni-hamburg.de/_media/research:theses:shree_krishna_adhikari_enhancing_openmp_for_parallelizing_input_output_in_shared_memory_applications.pdf}}},
	type	 = {Master's Thesis},
	abstract	 = {This thesis investigates the challenges of parallel I/O in shared memory applications,
			particularly within the OpenMP framework. It explores the limitations of traditional I/O
			approaches in OpenMP and proposes a novel technique to enhance parallel I/O performance. The
			proposed method involves careful management of thread offsets and data chunk distribution to
			enable efficient parallel write operations. Extensive experiments were conducted on diverse
			high-performance computing systems, utilizing various storage devices such as NVMe SSDs,
			CephFS, SATA SSDs, and HDDs. The results demonstrate that the proposed parallel I/O technique
			significantly improves performance on high-throughput storage devices, achieving throughput
			close to the theoretical maximum. While the gains were less pronounced on slower HDDs due to
			their mechanical limitations, the technique still exhibited improvements in specific
			scenarios. This research contributes to the advancement of parallel I/O capabilities in shared
			memory environments, offering a promising solution to overcome I/O bottlenecks and enhance the
			overall efficiency of OpenMP applications. The findings highlight the potential for further
			development and integration of parallel I/O techniques into the OpenMP framework, paving the
			way for improved performance and scalability in future high-performance computing
			applications.},
}