University of Houston

I/O is a major time-limiting factor in high performance computing (HPC) applications. The combined effects of hard drive latency and bandwidth make I/O the slowest operation in a system. A lot of work has been done in the field of parallel I/O for scientific computing, specially for distributed memory machines. As shared memory systems gain popularity with the increasing number of cores in a node, implementing efficient parallel I/O for shared memory machines has become an important challenge. Currently, popular shared memory programming models like OpenMP do not provide a framework for implementing parallel I/O. This thesis aims to develop parallel I/O for the shared memory architecture. In particular, focus has been laid on implementing parallel I/O for OpenMP. In the process, the characteristics of shared memory machines and the behaviour of parallel file systems has been studied and an effort has been made to optimize parallel I/O. Also, this research attempts at modeling parallel I/O, which enables us to predict behaviour of parallel I/O on systems with different configurations.