High Performance of Parallel I/O Methods in Supercomputers

Theme of bachelor thesis: «High Performance of Parallel I/O Methods in Supercomputers». The bachelor thesis includes the following sections: introduction, three chapters, conclusion, bibliography and two annexes. The introduction highlights the relevance of the topic in technical field of studies and demonstrates its practical significance. Moreover, much attention in introduction is given to subtasks, which should be successfully implemented to reach the main goal of the project. The first chapter covers the levels of I/O system organization, basics of parallel I/O hardware, RAID technology, distributed and parallel file systems, message passing interface (MPI-IO), high-level I/O libraries. Conclusion of the first chapter describes the problems in the implementation of parallel I/O. The second chapter draws our attention to the subtasks required to implement an efficient and user-friendly DVM I/O subsystem. It defines and demonstrates the methodology for performing mentioned subtasks: a set of necessary performance tests and a list of the considered computing systems with their characteristics. The end of the chapter presents conclusions and recommendations for different cluster architectures based on the undertaken test results on three cluster systems: «IBM Blue Gene/P», «Lomonosov», «MVS-100K». The third chapter gives a valuable information about the testing system and how it identifies the most productive method for considering cluster. Furthermore, it presents the way of organization for chosen I/O subsystem. The conclusion formulates the main findings from the experimental results of performance tests for parallel I/O technologies held on the IAM and MSU computer systems. This part also reports about the principle, which allows to analyze a particular architecture, determine and then realize the optimal method of parallel I/O.